From 73b2e63aac123f3ab9fb6e86314297b317fff53a Mon Sep 17 00:00:00 2001
From: Piotr Nawrot
Date: Thu, 16 Mar 2023 14:59:33 +0100
Subject: [PATCH] Release
---
LICENSE | 203 +++++++++++
README.md | 188 ++++++++++
assets/env_dump/lscpu.txt | 30 ++
assets/env_dump/nvidia_smi.txt | 20 ++
assets/env_dump/pip_freeze.txt | 154 +++++++++
assets/ft_loss.png | Bin 0 -> 96042 bytes
assets/ft_rougeL.png | Bin 0 -> 46095 bytes
assets/nanoT5.png | Bin 0 -> 415597 bytes
assets/pt_loss.png | Bin 0 -> 107278 bytes
nanoT5/__init__.py | 0
nanoT5/configs/default.yaml | 59 ++++
nanoT5/configs/task/ft.yaml | 31 ++
nanoT5/configs/task/pt.yaml | 1 +
nanoT5/main.py | 69 ++++
nanoT5/utils/__init__.py | 3 +
nanoT5/utils/copied_utils.py | 609 +++++++++++++++++++++++++++++++++
nanoT5/utils/gen_utils.py | 61 ++++
nanoT5/utils/logging_utils.py | 95 +++++
nanoT5/utils/model_utils.py | 321 +++++++++++++++++
nanoT5/utils/ni_dataset.py | 173 ++++++++++
nanoT5/utils/train_utils.py | 211 ++++++++++++
requirements.txt | 15 +
22 files changed, 2243 insertions(+)
create mode 100644 LICENSE
create mode 100644 README.md
create mode 100644 assets/env_dump/lscpu.txt
create mode 100644 assets/env_dump/nvidia_smi.txt
create mode 100644 assets/env_dump/pip_freeze.txt
create mode 100644 assets/ft_loss.png
create mode 100644 assets/ft_rougeL.png
create mode 100644 assets/nanoT5.png
create mode 100644 assets/pt_loss.png
create mode 100644 nanoT5/__init__.py
create mode 100644 nanoT5/configs/default.yaml
create mode 100644 nanoT5/configs/task/ft.yaml
create mode 100644 nanoT5/configs/task/pt.yaml
create mode 100644 nanoT5/main.py
create mode 100644 nanoT5/utils/__init__.py
create mode 100644 nanoT5/utils/copied_utils.py
create mode 100644 nanoT5/utils/gen_utils.py
create mode 100644 nanoT5/utils/logging_utils.py
create mode 100644 nanoT5/utils/model_utils.py
create mode 100644 nanoT5/utils/ni_dataset.py
create mode 100644 nanoT5/utils/train_utils.py
create mode 100644 requirements.txt
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000..ea4758d
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,203 @@
+Copyright 2022 - Piotr Nawrot. All rights reserved.
+
+ Apache License
+ Version 2.0, January 2004
+ http://www.apache.org/licenses/
+
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+ 1. Definitions.
+
+ "License" shall mean the terms and conditions for use, reproduction,
+ and distribution as defined by Sections 1 through 9 of this document.
+
+ "Licensor" shall mean the copyright owner or entity authorized by
+ the copyright owner that is granting the License.
+
+ "Legal Entity" shall mean the union of the acting entity and all
+ other entities that control, are controlled by, or are under common
+ control with that entity. For the purposes of this definition,
+ "control" means (i) the power, direct or indirect, to cause the
+ direction or management of such entity, whether by contract or
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
+ outstanding shares, or (iii) beneficial ownership of such entity.
+
+ "You" (or "Your") shall mean an individual or Legal Entity
+ exercising permissions granted by this License.
+
+ "Source" form shall mean the preferred form for making modifications,
+ including but not limited to software source code, documentation
+ source, and configuration files.
+
+ "Object" form shall mean any form resulting from mechanical
+ transformation or translation of a Source form, including but
+ not limited to compiled object code, generated documentation,
+ and conversions to other media types.
+
+ "Work" shall mean the work of authorship, whether in Source or
+ Object form, made available under the License, as indicated by a
+ copyright notice that is included in or attached to the work
+ (an example is provided in the Appendix below).
+
+ "Derivative Works" shall mean any work, whether in Source or Object
+ form, that is based on (or derived from) the Work and for which the
+ editorial revisions, annotations, elaborations, or other modifications
+ represent, as a whole, an original work of authorship. For the purposes
+ of this License, Derivative Works shall not include works that remain
+ separable from, or merely link (or bind by name) to the interfaces of,
+ the Work and Derivative Works thereof.
+
+ "Contribution" shall mean any work of authorship, including
+ the original version of the Work and any modifications or additions
+ to that Work or Derivative Works thereof, that is intentionally
+ submitted to Licensor for inclusion in the Work by the copyright owner
+ or by an individual or Legal Entity authorized to submit on behalf of
+ the copyright owner. For the purposes of this definition, "submitted"
+ means any form of electronic, verbal, or written communication sent
+ to the Licensor or its representatives, including but not limited to
+ communication on electronic mailing lists, source code control systems,
+ and issue tracking systems that are managed by, or on behalf of, the
+ Licensor for the purpose of discussing and improving the Work, but
+ excluding communication that is conspicuously marked or otherwise
+ designated in writing by the copyright owner as "Not a Contribution."
+
+ "Contributor" shall mean Licensor and any individual or Legal Entity
+ on behalf of whom a Contribution has been received by Licensor and
+ subsequently incorporated within the Work.
+
+ 2. Grant of Copyright License. Subject to the terms and conditions of
+ this License, each Contributor hereby grants to You a perpetual,
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+ copyright license to reproduce, prepare Derivative Works of,
+ publicly display, publicly perform, sublicense, and distribute the
+ Work and such Derivative Works in Source or Object form.
+
+ 3. Grant of Patent License. Subject to the terms and conditions of
+ this License, each Contributor hereby grants to You a perpetual,
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+ (except as stated in this section) patent license to make, have made,
+ use, offer to sell, sell, import, and otherwise transfer the Work,
+ where such license applies only to those patent claims licensable
+ by such Contributor that are necessarily infringed by their
+ Contribution(s) alone or by combination of their Contribution(s)
+ with the Work to which such Contribution(s) was submitted. If You
+ institute patent litigation against any entity (including a
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
+ or a Contribution incorporated within the Work constitutes direct
+ or contributory patent infringement, then any patent licenses
+ granted to You under this License for that Work shall terminate
+ as of the date such litigation is filed.
+
+ 4. Redistribution. You may reproduce and distribute copies of the
+ Work or Derivative Works thereof in any medium, with or without
+ modifications, and in Source or Object form, provided that You
+ meet the following conditions:
+
+ (a) You must give any other recipients of the Work or
+ Derivative Works a copy of this License; and
+
+ (b) You must cause any modified files to carry prominent notices
+ stating that You changed the files; and
+
+ (c) You must retain, in the Source form of any Derivative Works
+ that You distribute, all copyright, patent, trademark, and
+ attribution notices from the Source form of the Work,
+ excluding those notices that do not pertain to any part of
+ the Derivative Works; and
+
+ (d) If the Work includes a "NOTICE" text file as part of its
+ distribution, then any Derivative Works that You distribute must
+ include a readable copy of the attribution notices contained
+ within such NOTICE file, excluding those notices that do not
+ pertain to any part of the Derivative Works, in at least one
+ of the following places: within a NOTICE text file distributed
+ as part of the Derivative Works; within the Source form or
+ documentation, if provided along with the Derivative Works; or,
+ within a display generated by the Derivative Works, if and
+ wherever such third-party notices normally appear. The contents
+ of the NOTICE file are for informational purposes only and
+ do not modify the License. You may add Your own attribution
+ notices within Derivative Works that You distribute, alongside
+ or as an addendum to the NOTICE text from the Work, provided
+ that such additional attribution notices cannot be construed
+ as modifying the License.
+
+ You may add Your own copyright statement to Your modifications and
+ may provide additional or different license terms and conditions
+ for use, reproduction, or distribution of Your modifications, or
+ for any such Derivative Works as a whole, provided Your use,
+ reproduction, and distribution of the Work otherwise complies with
+ the conditions stated in this License.
+
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
+ any Contribution intentionally submitted for inclusion in the Work
+ by You to the Licensor shall be under the terms and conditions of
+ this License, without any additional terms or conditions.
+ Notwithstanding the above, nothing herein shall supersede or modify
+ the terms of any separate license agreement you may have executed
+ with Licensor regarding such Contributions.
+
+ 6. Trademarks. This License does not grant permission to use the trade
+ names, trademarks, service marks, or product names of the Licensor,
+ except as required for reasonable and customary use in describing the
+ origin of the Work and reproducing the content of the NOTICE file.
+
+ 7. Disclaimer of Warranty. Unless required by applicable law or
+ agreed to in writing, Licensor provides the Work (and each
+ Contributor provides its Contributions) on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ implied, including, without limitation, any warranties or conditions
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+ PARTICULAR PURPOSE. You are solely responsible for determining the
+ appropriateness of using or redistributing the Work and assume any
+ risks associated with Your exercise of permissions under this License.
+
+ 8. Limitation of Liability. In no event and under no legal theory,
+ whether in tort (including negligence), contract, or otherwise,
+ unless required by applicable law (such as deliberate and grossly
+ negligent acts) or agreed to in writing, shall any Contributor be
+ liable to You for damages, including any direct, indirect, special,
+ incidental, or consequential damages of any character arising as a
+ result of this License or out of the use or inability to use the
+ Work (including but not limited to damages for loss of goodwill,
+ work stoppage, computer failure or malfunction, or any and all
+ other commercial damages or losses), even if such Contributor
+ has been advised of the possibility of such damages.
+
+ 9. Accepting Warranty or Additional Liability. While redistributing
+ the Work or Derivative Works thereof, You may choose to offer,
+ and charge a fee for, acceptance of support, warranty, indemnity,
+ or other liability obligations and/or rights consistent with this
+ License. However, in accepting such obligations, You may act only
+ on Your own behalf and on Your sole responsibility, not on behalf
+ of any other Contributor, and only if You agree to indemnify,
+ defend, and hold each Contributor harmless for any liability
+ incurred by, or claims asserted against, such Contributor by reason
+ of your accepting any such warranty or additional liability.
+
+ END OF TERMS AND CONDITIONS
+
+ APPENDIX: How to apply the Apache License to your work.
+
+ To apply the Apache License to your work, attach the following
+ boilerplate notice, with the fields enclosed by brackets "[]"
+ replaced with your own identifying information. (Don't include
+ the brackets!) The text should be enclosed in the appropriate
+ comment syntax for the file format. We also recommend that a
+ file or class name and description of purpose be included on the
+ same "printed page" as the copyright notice for easier
+ identification within third-party archives.
+
+ Copyright [yyyy] [name of copyright owner]
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..9df59d2
--- /dev/null
+++ b/README.md
@@ -0,0 +1,188 @@
+# nanoT5 (Encoder-Decoder / Pre-training + Fine-Tuning)
+
+![nanoT5](assets/nanoT5.png)
+
+[**TLDR**](#tldr) | [**Motivation**](#motivation) | [**Setup**](#setup) | [**References**](#references) | [**Conclusions**](#conclusions) | [**Issues**](#issues)
+
+## TLDR:
+
+This repository contains the code to reproduce the pre-training of a "Large Language Model" (T5) under a limited budget (1xA100 GPU, ~20 hours) in PyTorch. We start from the randomly initialised T5-base-v1.1 (248M parameters) model implemented in HuggingFace. Next, we pre-train it on the English subset of the C4 dataset and then fine-tune it on Super-Natural Instructions (SNI).
+
+**In ~20 hours on a single GPU, we achieve ~40 RougeL on the SNI test set, compared to ~42 RougeL of the original model available on HuggingFace Hub and pretrained through "a combination of model and data parallelism [...] on slices of Cloud TPU Pods", each with 1024 TPUs.**
+
+Our core contribution is not the T5 model itself, which follows the HuggingFace implementation. Instead, we optimise everything else in the training pipeline to offer you a user-friendly starting template for your NLP application/research.
+
+## Motivation
+
+Despite the continuously increasing size of pretrained [Transformers](https://arxiv.org/pdf/1706.03762.pdf), the research community still needs easy-to-reproduce and up-to-date baselines to test new research hypotheses fast and at a small scale.
+
+A recent effort from Andrej Karpathy, the [nanoGPT](https://github.com/karpathy/nanoGPT) repository, enables researchers to pre-train and fine-tune GPT-style (Decoder-only) language models. On the other hand, [Cramming](https://github.com/JonasGeiping/cramming) implements the optimal BERT-style (Encoder-only) pre-training for limited-compute settings.
+
+With [nanoT5](https://github.com/PiotrNawrot/nanoT5), we want to fill a gap (Community requests: [#1](https://github.com/huggingface/transformers/issues/18030) [#2](https://github.com/facebookresearch/fairseq/issues/1899) [#3](https://github.com/google-research/text-to-text-transfer-transformer/issues/172) [#4](https://discuss.huggingface.co/t/example-of-how-to-pretrain-t5/4129) [#5](https://github.com/huggingface/transformers/issues/5079)) of an accessible research template to pre-train and fine-tune T5-style (Encoder-Decoder) model. **To the best of our knowledge, it is the first attempt to reproduce T5 v1.1 pre-training in PyTorch (previously available implementations are in Jax/Flax).**
+
+##
+
+**We created this repository for people who want to pre-train T5-style models by themselves and evaluate their performance on downstream tasks.** This could be for a variety of reasons:
+- You are a researcher in academia with limited compute (like me), and you came up with a promising idea to modify the T5 model, so you need a pipeline to evaluate it;
+- You have an in-house dataset that you think is more appropriate than the original pre-training data;
+- You want to experiment with continued pre-training or want to build on the T5 pre-training objective.
+
+**If you don't need to pre-train the T5 model, you'd be better off downloading the weights from HuggingFace Hub. Our checkpoints are worse because we work under limited compute.**
+
+##
+
+In this project, we expose (for research purposes) and optimise everything in the training pipeline of T5 except from model implementation. **Most importantly, we base our code on PyTorch, since access to TPUs is limited.** Among others:
+- **Dataset:** Downloading and preprocessing of the C4 dataset happens in parallel with the training of the model. The C4 dataset is > 300GB, so it takes a couple of hours to download it and even longer to preprocess it. This codebase does it on the fly without any detrimental effect on the training loss (we haven't observed it, although it might happen with an old CPU (< 8 core) or a slow internet connection). **As a result, you can start pre-training right after downloading and setting up this repository.**
+- **Model Optimizer / LR Scheduler:** The original T5 uses a memory-efficient Adafactor optimizer. [A study on pre-training T5](https://huggingface.co/spaces/yhavinga/pre-training-dutch-t5-models), on the other hand, reports that training does not converge with AdamW. We analysed the source of this discrepancy with several ablations. Although there are many subtle differences between Adafactor and AdamW, what ensures the Adafactor convergence is [matrix-wise LR scaling by its root mean square (RMS)](https://github.com/huggingface/transformers/blob/main/src/transformers/optimization.py#L595). We augmented the AdamW implementation by RMS scaling and observed that it becomes **more stable during pre-training, achieves better validation loss, and is faster**.
+- **Exposure and simplicity:** We try to balance the implementation of the training pipeline by keeping it customisable while retaining a sufficient level of abstraction. We use the [HuggingFace Accelerator](https://huggingface.co/docs/accelerate/index) to implement operations like Checkpoint Saving, Gradient Accumulation and moving tensors to the correct devices. We use [neptune.ai](https://neptune.ai) for experiment tracking and [hydra](https://hydra.cc/docs/intro/) for hyperparameter search. Apart from this, we expose the training loop, data preprocessing, etc.
+- **Efficiency:** We enable TF32 operations (Ampere GPUs) by default, use PyTorch 2.0 compile, and utilise all optimisations listed in established optimisation tutorials [#1](https://huggingface.co/docs/transformers/perf_train_gpu_one) [#2](https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html).
+
+## Setup
+
+### Environment & Hardware:
+
+```
+git clone https://github.com/PiotrNawrot/nanoT5.git
+cd nanoT5
+conda create -n nanoT5 python=3.8
+conda activate nanoT5
+pip3 install numpy --pre torch torchvision torchaudio --force-reinstall --index-url https://download.pytorch.org/whl/nightly/cu117
+pip install -r requirements.txt
+```
+
+The following commands result in the following [pip freeze](assets/env_dump/pip_freeze.txt) as of 15.03.2023.
+
+We also include our [lscpu](assets/env_dump/lscpu.txt) and [nvidia-smi](assets/env_dump/nvidia_smi.txt).
+
+### Pre-training:
+
+#### Reference:
+
+The [T5 v1.1](https://arxiv.org/pdf/2002.05202.pdf) authors report **1.942** negative log-likelihood (NLL) on the held-out set after after 2^16 steps.
+
+#### Legacy Optimizer (Adafactor) & LR Schedule (Inverse-Square-Root)
+
+We follow the original experimental setup for pre-training, including [Dataset (C4)](https://github.com/PiotrNawrot/nanoT5/blob/main/nanoT5/utils/model_utils.py#L58), [Training Objective (Span Filling)](https://github.com/PiotrNawrot/nanoT5/blob/main/nanoT5/utils/copied_utils.py#L16), [Model Architecture (T5-Base)](https://github.com/PiotrNawrot/nanoT5/blob/main/nanoT5/configs/default.yaml#L12), [Optimizer (Adafactor)](https://github.com/PiotrNawrot/nanoT5/blob/main/nanoT5/utils/model_utils.py#L236), and [LR Schedule (Inverse-Square-Root)](https://github.com/PiotrNawrot/nanoT5/blob/main/nanoT5/utils/model_utils.py#L276).
+
+Our negative log-likelihood on the held-out set is **1.995**, slightly worse than the reference.
+
+#### AdamW with RMS scaling Optimizer & Cosine LR Schedule
+
+We also experiment with the AdamW optimizer (instead of the original Adafactor) as it offers more stability during training. Instead of using a low-rank approximation for the second moment of the gradients, it estimates it directly by storing the moving average for each parameter in memory. However, training diverges with AdamW, similar to [this study on T5 pre-training](https://huggingface.co/spaces/yhavinga/pre-training-dutch-t5-models). Through several ablations, we found that [matrix-wise LR scaling by its root mean square (RMS)](https://github.com/huggingface/transformers/blob/main/src/transformers/optimization.py#L595) is responsible for the convergence of Adafactor. We augmented the AdamW implementation by RMS scaling and observed that [it converges, becomes more stable during pre-training](assets/pt_loss.png) and is slightly faster (it retrieves the second moment from memory instead of approximating it via matrix multiplications).
+
+However, AdamW, when paired with the Inverse-Square-Root LR schedule, performs worse than Adafactor. For our final experiment, we replace ISR with Cosine LR Schedule. We achieve **1.953** negative log-likelihood on the held-out set and significantly outperform Adafactor with ISR schedule.
+
+
+
+| | **Inverse-Square-Root** | **Cosine** |
+| :---: | :----: | :---: |
+| **Adafactor** | 1.995 | 1.993 |
+| **AdamW** | 2.040 | **1.953** |
+
+
+
+#### Increased BS (128 -> 144) to maximise GPU Utilization
+
+We notice that with the original Batch Size of 128, we use 60GB / 80GB GPU memory. To maximise the GPU Utilization by allowing for more parallelism, we increase the Batch Size to 144 and consider it **our default pre-training config**. This achieves **1.932** negative log-likelihood on the held-out set, improving upon all previous experiments.
+
+#### Training loss of experiments with different optimisers, schedulers, and batch sizes
+
+![pt_loss](assets/pt_loss.png)
+
+When not indicated in the plot, the batch size is 128.
+
+#### Examples
+
+To reproduce our default pre-training config experiment, run the following:
+
+```
+python -m nanoT5.main
+```
+
+To reproduce any of the experiments mentioned above choose any combination of hyperparameters as follows:
+
+```
+python -m nanoT5.main \
+ optim.name={adafactor,adamwscale} \
+ optim.batch_size={128,144} \
+ optim.lr_scheduler={legacy,cosine}
+```
+
+We recommend adding `model.compile=true` flag for pre-training, if you are able to install PyTorch 2.0. In our case it effects in 1.33x speedup.
+
+Suppose you don't have access to a 80GB GPU. In that case, you can increase the number of gradient accumulation steps by `optim.grad_acc=steps`, In where `batch_size` has to be divisible by `steps`.
+
+The summary of the optimization process is printed every 100 steps in the following format. For instance:
+
+```
+[train] Step 100 out of 65536 | Loss --> 59.881 | Grad_l2 --> 61.126 | Weights_l2 --> 7042.931 | Lr --> 0.010 | Seconds_per_step --> 1.385 |
+```
+
+### Fine-tuning:
+
+To fine-tune our model, we use the popular meta-dataset called **Super Natural-Instructions (SNI)**, which aggregates datasets for many tasks. This meta-datasets was used to fine-tune many of the recent LLMs, e.g. [FlanT5](https://arxiv.org/pdf/2210.11416.pdf), [BLOOM](https://arxiv.org/pdf/2211.05100.pdf), and [Tk-Instruct](https://arxiv.org/pdf/2204.07705.pdf). While FlanT5 and BLOOM use other corpora in addition to SNI, Tk-Instruct's pipeline consists of starting from a pre-trained T5 model and fine-tuning it solely on SNI.
+
+In this repository, we reproduce the Tk-Instruct fine-tuning results and use their pipeline to evaluate our pre-training config.
+
+#### Download the Super-Natural Instructions data:
+
+```
+git clone https://github.com/allenai/natural-instructions.git data
+```
+
+#### Run fine-tuning:
+
+We strictly follow the fine-tuning [config](nanoT5/configs/task/ft.yaml) of Tk-Instruct. It remains unclear whether Tk-Instruct was initialised from a regular checkpoint (*google/t5-v1_1-base*) or the one adapted explicitly for Language Modelling with continued training (*google/t5-base-lm-adapt*). Therefore, we decided to evaluate both. Run the following command to reproduce the Tk-Instruct experiments:
+
+```
+python -m adaptive.moe task=ft \
+ model.name={google/t5-v1_1-base,google/t5-base-lm-adapt} \
+ model.random_init={true,false} \
+ model.checkpoint_path={"","/path/to/pytorch_model.bin"}
+```
+
+Setting `model.random_init=false model.checkpoint_path=""` corresponds to downloading pre-trained weights from HuggingFace Hub.
+
+Setting `model.random_init=false model.checkpoint_path="/path/to/pytorch_model.bin"` corresponds to using the weights [**pre-trained**](#pre-training) with nanoT5.
+
+Setting `model.random_init=true model.checkpoint_path=""` corresponds to a random initialisation.
+
+
+#### Fine-tuning loss curves:
+
+![ft_loss](assets/ft_loss.png)
+
+#### Rouge-L on the held-out test-set:
+
+![ft_rougeL](assets/ft_rougeL.png)
+
+### Efficiency statistics:
+
+
+
+| | **Pre-training** | **Fine-tuning** |
+| :---: | :----: | :---: |
+| **One training step** | ~1.05s | ~0.175s |
+| **Steps** | 65536 | 18830 |
+| **Full training** | ~19h | ~1h |
+
+
+
+For pre-training we compile our model with PyTorch 2.0 using `model.compile=true` flag.
+
+## Conclusions:
+
+We show that it is possible to successfully pre-train a "Large Language Model" (T5) under a limited budget (1xA100 GPU, ~20 hours) in PyTorch. We make our codebase, configs and training logs publicly available to enhance the accessibility of NLP research. We are keen to hear your suggestions to improve the codebase further.
+
+## References:
+- [T5 paper](https://arxiv.org/pdf/1910.10683.pdf)
+- [T5 v1.1 paper](https://arxiv.org/pdf/2002.05202.pdf)
+- [Super-Natural Instructions paper](https://arxiv.org/pdf/2204.07705.pdf)
+- [HuggingFace Flax Script](https://github.com/huggingface/transformers/blob/main/examples/flax/language-modeling/run_t5_mlm_flax.py)
+- [Karpathy's nanoGPT](https://github.com/karpathy/nanoGPT)
+- [Instruct-GPT codebase (Super-Natural Instructions)](https://github.com/yizhongw/Tk-Instruct)
+- [Blog about pre-training Dutch T5 in HuggingFace](https://huggingface.co/spaces/yhavinga/pre-training-dutch-t5-models)
+
+## Issues:
+
+If you have any questions, feel free to raise a Github issue or contact me directly at: piotr.nawrot@ed.ac.uk
diff --git a/assets/env_dump/lscpu.txt b/assets/env_dump/lscpu.txt
new file mode 100644
index 0000000..78a4739
--- /dev/null
+++ b/assets/env_dump/lscpu.txt
@@ -0,0 +1,30 @@
+Architecture: x86_64
+CPU op-mode(s): 32-bit, 64-bit
+Byte Order: Little Endian
+CPU(s): 128
+On-line CPU(s) list: 0-127
+Thread(s) per core: 1
+Core(s) per socket: 64
+Socket(s): 2
+NUMA node(s): 8
+Vendor ID: AuthenticAMD
+CPU family: 25
+Model: 1
+Model name: AMD EPYC 7763 64-Core Processor
+Stepping: 1
+CPU MHz: 2445.206
+BogoMIPS: 4890.41
+Virtualization: AMD-V
+L1d cache: 32K
+L1i cache: 32K
+L2 cache: 512K
+L3 cache: 32768K
+NUMA node0 CPU(s): 0-15
+NUMA node1 CPU(s): 16-31
+NUMA node2 CPU(s): 32-47
+NUMA node3 CPU(s): 48-63
+NUMA node4 CPU(s): 64-79
+NUMA node5 CPU(s): 80-95
+NUMA node6 CPU(s): 96-111
+NUMA node7 CPU(s): 112-127
+Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca
diff --git a/assets/env_dump/nvidia_smi.txt b/assets/env_dump/nvidia_smi.txt
new file mode 100644
index 0000000..2f2112f
--- /dev/null
+++ b/assets/env_dump/nvidia_smi.txt
@@ -0,0 +1,20 @@
+Tue Mar 14 23:23:03 2023
++-----------------------------------------------------------------------------+
+| NVIDIA-SMI 510.85.02 Driver Version: 510.85.02 CUDA Version: 11.6 |
+|-------------------------------+----------------------+----------------------+
+| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
+| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
+| | | MIG M. |
+|===============================+======================+======================|
+| 0 NVIDIA A100-SXM... On | 00000000:01:00.0 Off | 0 |
+| N/A 55C P0 282W / 500W | 31620MiB / 81920MiB | 89% Default |
+| | | Disabled |
++-------------------------------+----------------------+----------------------+
+
++-----------------------------------------------------------------------------+
+| Processes: |
+| GPU GI CI PID Type Process name GPU Memory |
+| ID ID Usage |
+|=============================================================================|
+| 0 N/A N/A 963448 C python 31617MiB |
++-----------------------------------------------------------------------------+
diff --git a/assets/env_dump/pip_freeze.txt b/assets/env_dump/pip_freeze.txt
new file mode 100644
index 0000000..2800129
--- /dev/null
+++ b/assets/env_dump/pip_freeze.txt
@@ -0,0 +1,154 @@
+absl-py==1.4.0
+accelerate==0.17.1
+aiohttp==3.8.4
+aiosignal==1.3.1
+antlr4-python3-runtime==4.9.3
+anyio==3.6.2
+argon2-cffi==21.3.0
+argon2-cffi-bindings==21.2.0
+arrow==1.2.3
+asttokens==2.2.1
+async-timeout==4.0.2
+attrs==22.2.0
+backcall==0.2.0
+beautifulsoup4==4.11.2
+bleach==6.0.0
+boto3==1.26.91
+botocore==1.29.91
+bravado==11.0.3
+bravado-core==5.17.1
+certifi==2022.12.7
+cffi==1.15.1
+charset-normalizer==2.1.1
+click==8.1.3
+cmake==3.25.0
+comm==0.1.2
+datasets==2.10.1
+debugpy==1.6.6
+decorator==5.1.1
+defusedxml==0.7.1
+dill==0.3.6
+evaluate==0.4.0
+executing==1.2.0
+fancycompleter==0.9.1
+fastjsonschema==2.16.3
+filelock==3.9.0
+fqdn==1.5.1
+frozenlist==1.3.3
+fsspec==2023.3.0
+future==0.18.3
+gitdb==4.0.10
+GitPython==3.1.31
+huggingface-hub==0.13.2
+hydra-core==1.3.2
+idna==3.4
+importlib-metadata==6.0.0
+importlib-resources==5.12.0
+ipykernel==6.21.3
+ipython==8.11.0
+ipython-genutils==0.2.0
+isoduration==20.11.0
+jedi==0.18.2
+Jinja2==3.1.2
+jmespath==1.0.1
+joblib==1.2.0
+jsonpointer==2.3
+jsonref==1.1.0
+jsonschema==4.17.3
+jupyter-events==0.6.3
+jupyter_client==8.0.3
+jupyter_core==5.2.0
+jupyter_server==2.4.0
+jupyter_server_terminals==0.4.4
+jupyterlab-pygments==0.2.2
+lit==15.0.7
+MarkupSafe==2.1.2
+matplotlib-inline==0.1.6
+mistune==2.0.5
+monotonic==1.6
+mpmath==1.2.1
+msgpack==1.0.5
+multidict==6.0.4
+multiprocess==0.70.14
+nbclassic==0.5.3
+nbclient==0.7.2
+nbconvert==7.2.10
+nbformat==5.7.3
+neptune==1.0.2
+nest-asyncio==1.5.6
+networkx==3.0rc1
+nltk==3.8.1
+notebook==6.5.3
+notebook_shim==0.2.2
+numpy==1.24.1
+oauthlib==3.2.2
+omegaconf==2.3.0
+packaging==23.0
+pandas==1.5.3
+pandocfilters==1.5.0
+parso==0.8.3
+pdbpp==0.10.3
+pexpect==4.8.0
+pickleshare==0.7.5
+Pillow==9.3.0
+pkgutil_resolve_name==1.3.10
+platformdirs==3.1.1
+prometheus-client==0.16.0
+prompt-toolkit==3.0.38
+protobuf==3.20.3
+psutil==5.9.4
+ptyprocess==0.7.0
+pure-eval==0.2.2
+pyarrow==11.0.0
+pycparser==2.21
+Pygments==2.14.0
+PyJWT==2.6.0
+pynvml==11.5.0
+pyrepl==0.9.0
+pyrsistent==0.19.3
+python-dateutil==2.8.2
+python-json-logger==2.0.7
+pytorch-triton==2.1.0+2c32f43999
+pytz==2022.7.1
+PyYAML==6.0
+pyzmq==25.0.1
+regex==2022.10.31
+requests==2.28.1
+requests-oauthlib==1.3.1
+responses==0.18.0
+rfc3339-validator==0.1.4
+rfc3986-validator==0.1.1
+rfc3987==1.3.8
+rouge-score==0.1.2
+s3transfer==0.6.0
+Send2Trash==1.8.0
+sentencepiece==0.1.97
+simplejson==3.18.4
+six==1.16.0
+smmap==5.0.0
+sniffio==1.3.0
+soupsieve==2.4
+stack-data==0.6.2
+swagger-spec-validator==3.0.3
+sympy==1.11.1
+terminado==0.17.1
+tinycss2==1.2.1
+tokenizers==0.13.2
+torch==2.1.0.dev20230315+cu117
+torchaudio==2.0.0.dev20230313+cu117
+torchvision==0.15.0.dev20230315+cu117
+tornado==6.2
+tqdm==4.65.0
+traitlets==5.9.0
+transformers==4.27.0
+typing_extensions==4.4.0
+uri-template==1.2.0
+urllib3==1.26.13
+wcwidth==0.2.6
+webcolors==1.12
+webencodings==0.5.1
+websocket-client==1.5.1
+wmctrl==0.4
+xxhash==3.2.0
+yarl==1.8.2
+zipp==3.15.0
diff --git a/assets/ft_loss.png b/assets/ft_loss.png
new file mode 100644
index 0000000000000000000000000000000000000000..341feb129f05e887983b134bf44b823e7a8496e0
GIT binary patch
literal 96042
zcmb?@cRbha`}RjFO&XNUlu(N7P07ekS=oESj-K&lUwss~~mTXL%Osu!B8r$1j
z+wn3pTm07tn5=A#m>H%?b>Wwsw0`*14uc`mL0@=jLa8PgJPbxcXdKb
zF#EkFiVZJE<@V|$^4Gk)C@w21Weo=s2?+@YM@o2;FZXIW7#vEiNvq3K*Hlre2Jb6}PpujnjAlQ&g<-5KY2Zl0Ih)qkDy3Ly*3{J|>aj
z6m!LR88$)gjncXXKH*en#2%_Zcnc+B`Qr~w4r(uLf)5Dcs
zYS){DCr7LFqE_GraA#2^AFI^zzFeR3X
zWI8soPOrSr8DU}J
z6zuLAR;#z%jqB;tr_u?6aVb)f%P`v=6&`Lwo;wzpLi665vs?pJRnPG89!g8=e-F44
zK*19_-Vkhui*xGCRO2#ikI#&A8vh>1zdfDIztX~{D0p1EA&3(FHs8~iSuOh5S=rb!
z_5`W8t-VdUJ}r&CVPwdTj#Nxx@*>RJKlNhWM7e0w#VI&w{NRfXS%l9
zz4mK4u@GDlHaCxtRKMWgU98Qw=#TjLkv3MA
zKEZSQCMB=^$XwCz_Zj=;eyd6r&t2ys7aXOg=avCe>0j1Mw^gcSxCTY~3oTq`GmG+_
zE4TkJ9w(&u5ij6|;r86+&MsR%B`3H=Tl@0txYxmsR+Jl$-uE+xE$^#SA~l<8d{1ZW
z4Z&&@J4-w*MW>-JHNmq9WY#F)4KV=(vzo7
zCAn?P-&QYvn;P%xudL+}j)#97lVd;g_}aB=qynz&vZ-M^^QFNA)b4NLMNW#!H)(jm
z8dAX2{rzhm@3rqTw-qe7{cvZyQ;K;qJtJ!jUhr5|dwcMgFSljmcy>qE)s_%qL_{T6jtR8*tK#$ZDOwU>0MvO;rwwGF-
zTgR*e{|3Wf;W%WscnlWXT;JN-;$N)7pO~8J%rlb9Djqu}7C?GGl1U@qjpXv>hU>xn
z!hVF5Ilbm(TT?M*yUWd_XO44P^wqs*(R=m!wa$;XG%c|_VZ&Yrl-{tCp`oF+Z5vW>
zIIY7%?BL8?E@097e%GjtVyZo%eWEd}0O#TEUOeW{*z}j2np)1>{I*43{wNHdmX$U1
z{)^)_tJCrp3B7qna0u~Wvo`E3RLV(9pUKV5?WTGrn5SO0Iu0)w&2I8MJDd3;o%B^+
z-slex{Ko8gd0-P^GS!Ocn3%Fu8_b8wqG0h<^NsQD6)of&~+KDBHWNDOMxOD0JjY##HaAnQLH_R0e
zVddZio{iLWy{YB3SKv61UAf=%mgWIwbJUNT+u;u}wfiF8%2caf*C!biH&$2ePn|w3
zAuBuef=IAx#-7!BHDa)ad4p6b-J+6V!2Ah1c~*J<#%zsbnNWJ-T?t;
zboKNKJRU!Jl0kvpUYj-P{3sEH<8}B`_x#xDhYuenYE^pXn{>Uj8UNl?2G+fV0Jk@{lbL{NdoW>SVbN&>+;Lbo;{n4r-qZ(
z&cAYIW@d&oP^`s}O;AwKJ43zFWo1e_mRr96V1H);TTwBaRq8w$uDtmtBis3JZbR{U
zpIHV;7&`GW0kJ4p-xwPoHyQlq5Exvfn(1xEQ`iZ!b(4u{3X;e;9MJClck$VO&!V6s
z7DUkq=lxh!^$UXB{yBAzwQfa0h<~a@Zd@A--#U!zu-HgBp}luAU8!(NsVqw-FDYef
zo~`=%;la~F^D7#qj*|sg&2dN`^)vPnr7w5z1Fwq*y>k&15~^&nQ8I+>zqzxs0?~X0
zVmpiN#6ydiZqqwXR(P4W#vmQ&?C)+(+w)SnOq}1}89s=2;)kVUq6$5Hkpfz7?)N`_
zJe+EYdSZBV2R4<6Jn<*lxF(2f>3bz`-tcb4zOiH{YYrIWi)%Y|vAgdh!zqX;v^Q_Q
z4i?&f4+q*eyJFk_dmy>;7KO$)dmU_+1~DbganPJzCh_+uiH_m
zytc;{B|Tu@CF8zSfBEllWu>%M*oIhckL@;SI
z!MwCXka`?mbhvYA(ayCAv8As_xme6rBK1=$GO6FZ@KhN8h9HW2JL_}kkh*Wrq?T=t
zy_uMt)ZvUV=`Xx9Sn7n`FUI-oY;T7~MM;Q>z3kZ%5EP8>&QQ+tI`FL6nj+`49C&DG
zNLT8#0DE1#sGl()dDV{4FIoC8j(2LhBe^AxFS41bG~m^%6VY5&ZKJh*jjgSjs{*#v
zv-@tln@efwhtocnr(1uD%5zg^)m0!WDoVAJ?9wH%_ILr4p)wxK=}WhNihb}qbLDxj
zc~8@O=r>)6OJZsE%TN+_|zJyuYtgDAyq
zGv;$CyR>suI$1368muCd#)7JA6*z@yM|u*odD{QL4`^Xq_sI^EnH?dl^avBS9K;7na%H+~dl$7R+N
zkda*xtyB9l++_%|KYTXDc
zf>+8-(-~iJShy259lX{f7@>n*HX@*(m{TQ-GF;FCuFPbfKJ9Ufi@k#cu$V_4tFC@eg@-l}r%`=9QN
zlI12w0m^vv!24le9bpK<80+0l8-&ezvNWWWlrAkTEe$x22cpN4p6UD~hJLSV_6X
zr1V=#83mj*7stjvrBD0YVn|Zu&iNOdt2Ytn*v6dl&pe@MCCsto#@pF9n)4xVBKgm1
znqd5b+lSx&kId`Sj5`Q>^VgacyM5?&N`D3W_O;8A=7`
zH8XGCx;6GA&!Dhv!
z3KI*z(>)W+x96bGFDS!n8r&M#cSKu)_tAol#u;LbMz7)Ye=Fz9N3t2roVY1=;iLj=
zAF<}19xUa8v_lVDltvggV*=W5Q!3jU}4dMo>wdq2ezdD1^Fm8SshA)mt=mCGW41CBNT$w0isq9}z=d
ztAMSgfD382&Vvr+?2zO?_S!r)=j_xOiOVnVDot?H^y+oTs!w~4
zlL)gb@KSq`4ey+qK0DoFQ2LClR<~p8_}J~H+R)Hfa><i^iR5vdEsy
zg>0OTSL&HJ2>b)*_ZOL>BG3Uee0-%tSu^&$ETm^3D@f^L?==d1@uH!jf%Kr~W$KBH86wZ8z3QzSP
z6oh^e@BRGzu(J+$?(Xiixn_wIR6WdETG(H0umh8llG3sA=3~a
zbAKsUfOuXj+oH%I{>vhFY5}RZ=co1J#VWDbo@@4W!>2Z~wKQ4nnWR3Sgyn4&Xc%a3
zZ?6r3L|?z>VVc+5#>f*XaYFbNZ!+7KuD|(_d)51u5?V`wnCfy%Ox4r(Z;~0Z2W$#&
zW7A67xG)cl&!o2sS3e6}yJ&Av_C>AqF{>?;!5kz3os8hejx=>!pEDPp98N&f!0+x8
zOyk&Kn@c#oAPb4#-u)c}FY!o%_k;Hz3hFZPKR=tQ
zUv3vmNL(KB<9u-9^jXzNCpEDx7u>whAn7Hkd#U>vV#=YQACfaWIQ~X$=Q}>=n%P{zm12yn=#HLIO3^6tIP$jwL!rLGgEH(PxqSp8=jDlX?{V
zA&ro)I+Ca>9xK+yDA^_@)^=8h1+(e&7bn@~bTCsgG0>_uwcmL7Ze(vC_z|*L4sb(;
z_zvZF0Zs_&MN}97E_-NipHFa|xMt}>NBIf2O!(>{xb1pAYl_MO32!ZxElwB?^};gI@tRuH#--J#vq6IC9l%Bic16
z-zW0yBgPvC9#V&!QysF*Z{6%tAtk6G8uSRlB{$yr!4XJeCP`VDH8
zb0j1w0z%G<+Gq@z=PzEEbf=S;n3(7r8Un7#K6mb16(DYrRM|LEPIFQy%_jV)J=12S
zF4m0e!=pye`ww}u)NvGjv2Sq9eBx%@OFgo0k$sQleNGB}$nw@3?o-X#!J3f@r6d?p
zE2yNE#6nVjLBL|a8y@(L{7lx8aCyZt(NLgRazrr^CGR`Wz3mQQ-vDVl1K)&7#PjFR
z2avKPt>a%#ivRJWry2ZfU`Fb79e(w4P{armSt;#
zqKaZLKy0JJUu$+gjgWk%85dgWA{y6V+V0f4&_&?WvsI8@qtTNkiFWx3W+Qg8D4=$M
zOB{z{;7;CLmz>ZIC_SUtjE7F&|u(qqE(y3@?xiA{S*h-kKbDums`!03t3w6o3K=st$wz4RD-j)
zbI+`HCd6*xVOzmk&ka<|c3|vhI(>n?Wd(){p=oDl?%v*BK(R}cky;l>r%jLSY>o%7
z{MHN3wZqBAyUp`J_4QCz7U)|6cc^SIs4N9u<@~vG_dGoXVIKLo2JJ6GEOq7_?;!>K%X8n6U4`%!uqajz)ASgE~KR?kY|JGx>
z#O`*oQpeKT;~pwfB_nJaQ|1>9G$@$v)6Wl|_-?3yy4)*gT?IED+{TsI)KDruMCo#c
z+af|WxZS}L+c!J(KaJqR;RtZDIn@OLst(avGkZrmwu(bt+>|6U8;++m_dWPCvCHvc
zTEkZ^z13g!t6&+L2_NHP1d^Q9XC?+71#V5H20KyeWhT4e95ZQRH{KPHKff*?V5m^q
z&7q}*rdpWY7?p_pZG*X#>eq4If)%tpMc9rqJ1g@Osne_P^;9w1JT5E9d9@ksRA;HL3u
z2%m4$jFX)+qIz;!W;B-bblcdGs?tz93?Nzm;D!27_g7U+$}YSV@TjC%K-X+u`|h5b
zqXeJ>fXs0Q2!l^g^9-eGNbk5V$6DT3Y*{>7EL5F7Yd3eBObjOraJDbcT&W4<8dAyB3#JxogV7RI*|iHO&v*ja
zpO(h3K0l~7&Ry<{C`JCPzp-`;l5K+;6C~y5#5;sOGF%!*h8OzBSAbd
z1`Q3wiQ$$6p#%tg%RjCM|Jq)IyfpJWS1&UG+rPQB)lqCSZb}*U_U*k4rOf;WyQ!9&
zc^<%IK&}sjQvWUb0ySZ}?7|9AujZWOfB_yTcAHYp+VcYGYX~Zi%55X7O*3|~nW1vG
z!0}Z&0B?Y3^o#?{_Va5sKh>50X~62(#^W@_?&PZm{`vEQCySOxPT;v|(Ff81s-TBZ
zk83QOT1-=RL2;CK8%@4M@F@4iK321qVGz>|Y2)kZ&I
zYya(y-fr%Ujlk-nyeEj^nfKDM(6bFJ6
zbOV1@YPwjngVWDYNfrF$jF_(2-sx`32MgpXbK_Fk;q&>#No+gas^*T&u~zejCti!;
z$i4FzsHOhMyN{)>;{!t|Qu{IeR$~JHiyJ1D4Z3%^EX4;j*xGjRX9ot_Fbvzqz~uVUf*cSzjqjy%A)c>CTkX2o(XL
z!^^1jB~kRvrIr}3U~g~#NY_2k3Za&-G0a=)HA;YG*F)MUFj2S43wIo>dcXQkHb;&x
zEVKiGQal-<*3@zx&7qEt4irCm=Dn;yz}~O)+#}_(R4?KFqQrRmGSkPLoVP#R1S*
zg3MvoKX_vfv{+q0Wbi|5Albbnr1Vcu|Jzsa2IUQq8i1N7<+V3X*{}lsR63gdWJY%r
zQcXZ*LNY)w|MH20%?2TmI4cXYE1na%q^Eyti05_iFP;DP1{iJ@kQY#9+TGjZv7I;p
zJoqP2&~{-)5e3M%Pm1yo$1h4(Z%%Y(t
zkpDEtv5!em#3_4j)g{XzsHFT=>09zu0yV=^Q@<}2T#9@!`)l6gkRH)flec#!kC=FX
zK3*qUU7B8Os@t8e2+AVS?l|?q&!$f+D=S;r%KiGLaK_hxfd(jRuJ;;A$MJ>(T^C`L
z;1N-|zx@W(6`Ns?4V!L75l^4>oDsLL4X@1)pkd2DT<4kY6*Ig@H`D!smD!
zWQUJ%j2tdi!uf`TN5ftSAxF>%NPrv?GuI1HUp(6*?MdkdB+BTBSg8htW=C%N
z3&Sz~lmBH9i1}QVyD_aKL8?6;zYJPbKR@<&pd>6=Jou~isPH$(xpa%uCr*q2he^-L
z35TZeZFsmS#C8l_*z2mQds$`6l5-_M;AhVCgA@v+?AuFN5ZaRBoM4c}BzFk}X89x&mj#vdt$g907@`?m_Pj)r+G
z+Ep0fKQY*8YenADhWh#?h!A?LQSbSjEK?eJs*tXZ(&<-&X5<|Z+FiuIE*8Lz*0W;jz8?d3E@{MafmC`eC*gU^WLv11CDvU
zEqU#NZ)!&q9cFtBqF=?iXuPQQ?wUEO4X1K*M|mK>JFewhX)%7^Ln}#1N#AT-3(*aG
zUX7HqtuTQyy1LisWn+J4;bb4RJIwj`90krPY6T8uKp4G7WizspK;|}itq%Y>rO^tc
z93CE?07||H6s3@c>@=?~7EE=v>h(o><6mj=z70wfAoyLsdGn+eJEOUFILvs;
z*9$w6#q(f8VF)R>??DtycMpt?=5;WQa^|%Bvbj+L`b=ol%(eg%NXH25pm+;Db-MS{kX@G64R7h#OoQjBr?=
z;9m)@Pjh6%K>fsnbCHWI@Qc;*cBqSs?eCNu$tQNL-J2q-d}Q!!S^;{axZ&v%2~zCA}Qijw=S
zm0ghhc-%Md-cePsYD&<|#x@o4f#l0I?4G>mvf3eDuqPHsCJM{8f@{z2vR&w;@>uO)
zU9$h9u=MLkY%E1zzDY7=*M|oL@G-?tWO(Y?4sf0ktpJ)v*wzZ}A|+)jh)3Xb$S6#x
z9Zg$&kkUsB{N>p5#K2f1TIFxS9BT|)e4bdU%QwtxBRtlZY5gv9`QT&Kojo2MP1Ke6
zT>MK0Q0XGAUoBo@p#8AxY~}!b@$YYrQwX>cTi?im%?XyDx3RG;Ag(ch1Pru#A5R-a
zxuXyj@Bzb~J9qkvtO!N`oOx&PsNmY-08#OQ$K$i(0L0)ca0u_;zu(7mi_gg*c+hcv
zpb^%j6^`ix*ft;|;n&yKqY`iO3$|jVjZXp&;$%~J9|!E9@0&~Gc$-k=q0?FAjc5Po
z*K3Hmx&^AubrV5S*;u>Uvs`!K5DV^a>Ty~PwJ>IvJqM-=K!i3h!UkiUQPQB)j05I_
z^Z>e^`R^P&Oib%{C}>+N5B97u?j9Z{aPrpaZ{3pQlLNN`Ea4~!gA;vb*_n$1wv$GrW$b~L%)oMw<5xIa{=
zb+U9hNHRD#@^*^hiPbh99A@fw1o6}-Jqf{AR$ZsPER!!JRYS
z)Ly$GkWaqajEjNGVGAN?3H*Mv`vEN0LEv`ZT(rIZ6ZAzO|L+>Mkb;H6nw#RiG#!-C
zH(nA^Yt9r@b2PZO#q#uhH3atR=K~^>>pvlxe0iGp668*$|@q?|ZJVO->!LYcv*w!rRGB7-_O+sL*;5jD46#@_!(v~aZz;X6qr^;aM
zPdCn-oy%j}eyT058HNKn7sA9d&2m?~yEbDY2sdz=|E}^mc~J#GGt3UllTXs4kmG1U
z2nNg46Ls|#aQ+;rLSYKZrUux&!cd?=6?2`A?o-ne)+ecFRWIPUF&rt3{ro_0ZOA)C
z0D@4jz;HaGCTT@#X*PbY#P;Y0m2n>T{&cXJHV|s=`?_D^=_>#9nYx2vuiRKk+O^q(
zM4}TZjkODp^>ZUi{nBg2#+tKARAuNrUJWs)d1@(;7-a87w+su_S9&lfSw;VyvE~=6
zI25#yw$ukBm4+DA8pFlhW&I@!v<&|@Z$QxG$-a@pbO~@Nk4!XsN9euOE^rMThfszM
zU5s{JY2~`YWBW0JmXCrkG%W1tSAB^nrdgvzf5yYK4i8ui_hoXhf~)4~wWVII&M@*%;)3uZ7Np%<88dt9Edj`BPPez;hh
zv{u(ui&jxqeWzd-$W^wl9fkU!tL{6nQ7Nhb^l9GmL9
zFC*;D+xaC%Kir`$7rG+V>1l5$lqAL<&}Pi!B&}U!hGFPiBjBbqC^RvPb+buzGwb|a**5qE`(jce5wxkR1{w&8ery87w+7c2$l
zy0bspkCHY0C}HF&ise@rFA(|Xci<4~2}>)SjCV{yS-C5*IV-tikd5V~QUzT~64yEB
z#(S8ShN6gG$8TQ*KFW?|%6ytOOzaxJ-BKT~=DGh5`?F|?iX|pe-WMl_r{XjW{&MZV
zcr`o!fQ7W(MO_7hL(YgDWWGP!$d&k;~=@
z5l=m`w!SWCV}T(drgS|>w4+)$D@%_P7od0z3pDx_Jj6&2$^|9kdk|%FntZbH;5QBs
zJ-7Yk!vh0ArQh{w-
z;=`)$GqUnzVj&?nzn&aX9Uhcd_a15(u04;uufHAs{tki2ABPHZaD8@udCGv7yR*C2y*x3K|kmh;j2?Cp&K&I}AlBH_WvFah4sq?QT&lr34
z$Aq-fMpJ(vc~Z%*8+M6}O)lsBDF)&jmJOl#M8>48BLl0T2fet4$w`qgK}_yKVpqvy
zLU7pfFL?eP@|?n^1X2rC1SMsZuxYiIpc)QTDrx}U2m%pO3l3ffJ68^vo4(o42t%PU
zGg!LeLXQh1
z`RJ1_S(V^Xs!Ubo*`^Y|v)$$Sm
zBv7q=H%G65P;loj8}iD5rM(5>e)0UbSqwsTU{L}t3UcUhjzfl-N&k~n*ZiwyXMR%(N_NUzE)@~-`wF+1Zzb;gLSC;kZChi6}pl}M|{J)I+|{!
z0@)>!({FWoxy9{WiB}6xwh)nur`{9^ys(NY31a(i!BU!t0`c`py0(_>gsm%3Ofevh9-LvKbHLVm1GL7q;DF9bLjKLo1jU8fYE>W}MaW
zuL`f@--lq5Rw=TKi!djm;C>5GYySg`@w2zIhAJUZEt=VXdNYdCTc1$Z6R0Ger#sf78tz{D0m9A
zA573^`$_<=BeSy^0+!e;%BN4Wh<+bzjuQc>2Ez>{;%VfQ#d2Y25Ly9jr2<#nM{pKM
z2^wHFpf4eUuPBKD+6I*PHlkQrS%2r7T!dvDhh_s^ATL*Ey8J+IuHxtv(5~{fs}KhY
zXaWw2Av9zFg}3l!X9()QSfDO}^*iltRk8IFOsZS4*fc5NQ9wD40OHIJR&NQsw3xeq
zf3qsKO`#U~4qRSidpkju_Fe1c2?-=g22l%T$!H;0AsD)i0E)ZA89+oHylq%$=vXdR
ziCTYR;PjQXtcihJV4e8ICu(&yXZGsXgY8q+96n|brRB3qbb9jj($a6)Rlx>6lEiST
zoya|DzPazs8#zbxvN~pfk=85f2OkavpJM<}AsIE?va#IHz*Yp?yw48U
zJ4}^jkzQW!t+`H5I#GF2SUfn7$J8Y@AOz*l5%Di#hy*vULemZEGk_&PHh(y(Cr+QP
z1&wxwm%{YyvwQ1#U#F
z1_)2z14cim#v>=|$jAum02rHtXKnxsV
z&s0D|L&K9tkM2Ew9ICW^7O|TuD(^YvcBJRG`d}zJP4CXaJ-UVo_dZW(wVN+lE+g#93L3|p$^8GEt^KV*yH0;L{Heasc
z-|HY}5iRz7of{wOf}~3Rc7af{qDV?-+hEg3su>3_jqZ+NWc-GMa{LVGJ+M9)_vPPI
z%E?O`{D`Qb9$~?_Ln{x|tDWG~
z2I5uKQbAQwQ9<#TdZP<4#?BDKHIu1c@11DHC=U{@mIl8-yip
zs?%d9E3iPkh1(XKS7hFK8x~dzlo5~@NQi`@4wZM0A720}5FJktnvf9(UfVl89Vzdd
zo6hL{Dh~GD5aADUPsV^i#fQ)kJ7)b~QedXwzoWZ&RL<{_-lsG}l_$YEGJG$+VyNN?
z#X(XQHHSQn?CjndGy?#jWPl>)v7bJd;J$nkqvgJIth%~-+Ws5s4g`?(*yNA@mueri
z%pi9Rl2GD3w@twR1-(BL@V|J8kci7xJEc%l7QA8X*qAQzu|Qwa(RW*rD
zS|<-BAprzbMnKEl6of$5YM9vZS_j6<0fP>W5na%F0;D0-jfoy^*d`V@iiSWQ0ZvPR
zaJvE`2<()~yY8FTsG{xc>H=o{EbJI{ue~*B7D8l0yvO=eAjiZ&5Je&jQ1OcXl^-7e
ziuMfX#LNwpTuhXH`ihI;_HBLei|1PnQwa$PO_^`{D88O8PHsLs@Y6FVwf3T-^^^ed
z*>+!hqC70qjg<}GO{eJxtAPiL_S|bz!r%NZGyY>Z{hvsI>}B=IK5%maH~2j@lm|C-
z*bg=FBZS>*h{C!)0hq(k8&dhQ(;RDHida0Pb-Eo(;{eL%nsm`+$DbbjIX~(JId>(jGfP1>UtJ+^<2Lj0H67m}4I(rpuO^+)K&{rruTd&REa2|fANbN;1semjb
zhZ^X>>T}}kS>zuc@H*JTY254w6B4wH5u!{6S>pr|Q7R0XAM|L|UbhDvpnn5)R-ai%
zx)n4}x&}NfmO!xh>$~W4WBk+jwLiCxj@)i*h>rK+`BZWkjA{O$p051kmIaO4chO8N
z!B_WoiGame-RpS5`wGG+Y<+-X&%Xs#Tx9L0zpZQe3Y%nqT!ip)VRoHKb}5JLVW}&!
zh8~t5VwZ~Kb<^*}PeWrEjK>UZIGW;#%SP3u>)NQ4w-Bj=B
z&>*Jzed6BGdU)EhSHNNIjs5SB@W#bbVX`Scl~>VtFp2t4VDY{{xqKZXOuTXW8G*hCe$Tm
zea~@xWZTg_@%l)W?by|*48dbp^EYICRx8eXC)M`n8l<}37DLZCVG~V@)BHgT%l=Ar
zgvmngU=sh(`gZ-~H+RzSKS{xucU%D`W7Yia;p-iNu`^ylR+*$r@2_bzmIyE&^Uleb
z7cshv7t3TZTydd$W>MN3EjWHSbYOyjQ34IC=?(+XumhV`aADR8&@I5~1Ze!jm9;Bg
zgk~Oh$W3R+*k3_wfor9wtt~5a*`jd4u$Pq)^w}oyhiM{ZVv9|btGj};q2PPv1M0b4
zuYSg#TgOL!G$&3ka_*{XOua_4pgscawXFyH8$w{ws)js?C~>G#=xAu3LEsO7&Iaqb
z0_?Ka!G;%uf6|;=^PQq|#KfRDPJo-7mX7YLZOVp%o2a-rpp=Yo<9>5aXc>jZE)q*{
zQzKFb1V_xzTs?6F+)r(sB_cvHg*_zMzFh3_7u5*G3xvMMpI&qA?Cdn1d_ib+?K(_x
zTBaTHtGg7&4q#mai$bmw8JbJ$m%q<;%*MddRp1aR*QN
znjlOzYkws)BSWpjB)Gfy0DTRC^Fy+|8)|LnlgprXzQaNm2=QY;4-4hSCvXH5I97m1
z{8wK|n;+U9gPDU&gTNo+s)?z@r^B7jdCobzlNB)*38c+l#hfCttcg!h93Djl{ljVd
ztar()esw9-r9}x_4knPavy-Py=UcFLOsMb!w`PN7(F!P2F|~d~X8ja|1aB@ge*z^I
zslh>RC}-*X)k`Q?cI%{Sju9{3x@4-ytL1E{^%8*@keccv}kiHEleFcLvmF
zIG^ZWYzges<;hJbv#31(GnaY&uYx(xXYMAN$#)SEyxc+
zx~B=fPlZy)^{Qt}M5U)AoYlwRCi?$Y4x79l{67}M7yu^Fufho#-v~qjM5+V-$YMYJ
z2s-$lG6jME0~MoS-rj=Nj}qYe3naZeZ_I(|M;|_qX^G}|gJ=x*dmx;%iWk0OHnkUlNVxv}l49e*y}3XyO8l0iF;fgFOJM
z9>D-a*8?7|ow5H4JOuQDJ}odML0zjT(b9#@#M*pJO$4YM(|d63v5`h|D>XJ`98aXmN`IbHo4U6_Er~)4t_8
z>km8gVZ`H$!oMP`vR>rG{+8>ZP6%0RFfu6QS!O#)t+2)X0K(9pNbr`#E}wDPS$Dhr
zvTA*9dbcxx?8Y<{bMy=h58YxTBb8?=5GL8*-v_Y-^zmlN4K=8rW4NrQAh{8!+YKZq
zCtu;Rq<}n%#B7kOI+bWh6XPT
zxWRLvrU7CE-2nje#0<@Q$V6e}Ay)+5^bip`d9UW1-=6cPyshDss&*az=9$Z
z2{4ILKoB|u%TfY*@N{1RBWkGzN5orc`q8)|~9n+(r?Vv&hDVYNxf67U%*p>?1gwEYp}g^e3BuDA{@12BFzJRA-o7&;8?U-)(4~Z
zc#s$GmtKyC)TR03V;C|yuU$`kK)w1*2gBfc9XNUEjw~Mi`i6B^R`$U3bQLz)#3q6W
z;S+dNOg99nAuJhzv7|2{&!I1#--VUAa+meZr8`Mb-h|H=0apZKWkCa+t?P`<9E!70
z7+mGx2zQ)Qe^Ls!I>1@f1#Or+P8QU2xnL@Rd?GSHkPo0o5bB28PL`tb#$dc~oM$u8
z*Utgm1vC2{I08t&VXP7E?Z83reyLQlQ){dS9
zI{@s=%HiZ$R|3VrG>wP6%uslPZSO3=b67H#C}~i6PkQI)=S#M)06MY(D^LhDKEgK4
ze}5W+6O_g!a9#^Mkn6x>u-wf_6X;^t`IQIze1PnCcio|pb98hRGho|%2Y4*d);7*j
zQ@6u`Qcb}sB^)vilHUsWVhLkqQyXw$2wn1q5iM<026jDuC3y9yXopU-*L=>!L#J?@R6B;Ik--dH;xI9C5a(XMehn`6Bi47E4D-VGw2OdSgsL|L>^I;)f~sS(FoECYf;aG2
zc@P63#k-crL*})C5?GP6;lE50vTkhxW9CKuxxMD$jwV83Nerl86qwN?i@+A%g-azz|%hxCCo|K2}XWc3$4o>N+4w
z)*0F1OEFzRPjS6k0iX$Tf+iR@v&wwtK_1=W0;>~`9OnIl
zT{MJL{AnI1Y^F+}z-)sLD5RJHCJc6r$qqQv=t{E@Yln$p_k}FPCMtiI%QpPnC3?z5
zYj*8sNxZ7~$eGcHh>G{A9mS|$|B+H6UNG1%<)yJ*k_V^!e{ouvuDm5UOC}OVCzH?W
zfVcw0tU{d%g|`mq)PMVnh^#rh`9R}`&4Gu!dUU`|KxXZPdKO$&O?4^h>72ael$0|6
z?~F%2ogsn+Ba$!G3e7+7ojiTItL=%3N}9(WU;w16qE!M<5ws-NS{CI&*Y#W(x;O$h
zejIgpm
z7O0IU2??h_9b*QH1zif{wZD=8Gz-k(AvT6vBj92K0|=9`+&0g_ZTJp=V!Cwd7St(A
zKozZk-DVjvSvVKG!&*U*=Kd(jJpD$_XqemYMdueSeu71;5{^?FHx?l!g=rNjzn!^d}Veqzh7S=a9DPyRE8AsNLSYsB#;SZkZGm&eU-83XnIdVJv>N
z)o_nVv;A7uYh>&Pd>HcX-A4f&Q$fX{m^F~(siHeK4)@X&j|cHNLoL7KplhnHA8P<(aZ|6+)>P>!R{6vzjbLD3tZWe9nA
z^M=-zoy2fyfQSjiyAHDAZn_uMRS&@61z!rYxJzrPaEyTYO|D59E(>{A;C9pLX=&?C
z)DP1RR{PW5HyZ6^Mw{qyh`L8TEn!eGG<{@*mLk;ZN2A)X{!S#;F+*I}+R0S;OqONF
ze?F-Ag27BqQc@Qj^&D
zy9rS-S!I8l8e|1LcG#+9z6zj|7VqzLSWcM7a_zGX2@R2e`$T#+d+}%HnA(R#|38H
zPcQz6u)M$9P;lg40ZFpzDS7wZSph2OcDDi?hEaX4Ne$W>m=nN!D~rAAZ(e@|o>~dU
z&bgg6clK8k?r_@^v;ZRdV|!zv9U3H5$K`GeBv?9!0X5C)g+Hx;rB6O4>Z^GN
zje+b^i64pG%Q|RQ4T+Q(`aBNw;Otyq#Jgk+DEBuDGZ+ko6r^ajOd3gRnzLCL|NAU^
z07DKeigpoJj>Evlh6PR%h(<1Ju0NA(ZHZzeI3q&GYYIY2?*pYF8lC_430fk+
zomq>;UiMz=(Rl0h_U9|>z9N%_(G&G03id>zwbwRsY)k0*#70I_*}uRCyJQCrKU6f6d;1?q(=&ha8rkpMa&uL9+g$Sssn9`0|F5
z7j*@H1%Ck94cy`F49fjFJ*VKth&!BGc8y%2gxP`0@07Qx{9j7U>$n7cqP#h8%vy6%Xz~3k;a%z>NRI
zO;(Gwyi-%zNAc|OV+nh3TqU<^Zz1?njM|}V6+$p{^FdT^qg1e25+DG#sKWZg_+_B
z3O_)d9mGOOJd$=xp7=C&G-f?}*Rb>be+OHP`MLtc>8w?v(pGzZ4yOi|bzNl+P+Ou4
z&v;8E0D7=Ao#04E|NHV!il33V?{jI6mi)GQn25E6Y4j6UF%Xy@;4vaka7RyhT?u*0
zCMtIjx#!5GFU^?Thut0TLF#NW=S#NR>oUIIbT^y^d~2s9R!${5cv8YviQWg_|EJz{
z)Z`D*gEr*bF4mWY2w6@2e!
zuI>M(F;->Ao1soDNpt?wXwFt)F{%GEp9PzZTSv{4`PWEx0NoZOwwJ#f5qsMw)OtVE
z?!`Bru=fcvZ@9K;Bi|p#?(Hdnw|+Eo_*LnF5M5lQ%FM~~2egu-D8k?Va2IbAi)byk
zblk_qhu_-zgM=Ci&flNi=Lve|S7w5uTio4!Vv>tIz1wl`VONzzX?;^gY^G5iyX3Uq
zjjm9`9>0v|_H1{-?k$o*hyupTX`j{iJY4L;x-amP@&ISC4g=uBI+qI`V&++~O(#oZU2%xFjKRNNLmbNx_?pI}9aJ-6KqwiMXqT;XF%-LVQB9c)f
zvL7>IEJpCx;PYbYarMg1RQdwWZH9A({7#g}I_kQ_2nWZG7S#uYqQIffUxAHDxvIf@YIETUrha#q
ze0;Fr%=uKL`5ExX1lh6*X!Sfd!J~MPw&rZ(DW#Vr`RHSEtobW@lR2RSf){cArx@g2
zmV^4#w
zDEg0IlH_zHUDx#iH*`zsQfL5862H%IX!azgfZ~6e-w|cl(rE(emyn+j&>-!cJ8&6A
z>bhu{y7cFOU}QYFTDTY7X<+0*^zS$IN4^fwcOoDG4QMN7t}P*RD@#*pBy(L1lsQjl
zIIVom_aIU57Q$Q$$PhxliWCGYghPPTe>C%xU+|`pO&9B8r0`84K(rbSs62iU&)w
z3s+T+_T>D#$)PtT%>zg53w3R}INbr0ZnTTos72iIIq&Nz>*m8i0shpUiGM%hIQ)pd
zxtETJ&iA(UJ(9-LuKjMLt|j>cjW0LKs1mj|GI}m(MO8lKWA{gXh1wJUIWN_#d08Lh
z4xl}Ie6}qR(F0pv#q8dNZ9V_8v3;A=Od?utdil-#`CI?*8!yejqb7ROC{Uo@mp(1;)FA`wMcUa8`FoFnPEhet
znH!ZkzG-wX;7)AprTWBu
z8>7%E+cQOCw$U%-gB2@~u=J|*6GQOFlvolVD18~pwY{U^42GV$c8UOy#9wmgo%brM
zJX(ig@Z!~&YTk`<-2u}uJz%w3R|K-cD}~}n}jklp1^eLPcXm_
zI@bOL_Hcc$r2Y(Si}@Ju5zjlZwf*_p5+!C#(N#Wb2YA^mS2bT{XSTK^v75$uibC7>
zRW%?g_b}C;hYxu5QUh6@A=sk@)Wx{@qhw0wf_iMKL8D~QiT}2rTnc^m2$h{4id|)}
z9T0kC-FVZXIs|mu;fe*HGCoHg3Gd0d49_^*cLV)9
zpNepO{n?mF&75gk@e=mDCaTwH#ieOA2amHf)jS
zar-;cKfewkY7n{aam)uN^K{31LLHf2hNDJ0*XK_`}6{OW9nryw9p-et|Lr
zt1lBBi6uA4&MPUl!aIKTCLGyMpF{$V@UI!F-EkiA6&tpYc91G4t=&01pJ*aTk87f{28hPcnu=KEy%v9(!#27#AkHy9@-;vCcp7D7)32lVSl~jue#EGAw
zjbWAWpqqSq&`d$MXhI^lg`4&A%MNj%0r
zu~j&$(mrPQ>9QAFh2Jq>JFnlP)6P#~JrsGjt96f3MdY}Us~nh+
zZwv)H4`MJ}4vot!hodREbw=%{B!9a@;+jq{53)L`BeQIT>FzWh7~ITtO!zTIbEYj*
zR`OedEnzjM>g~$5n);J1;jFCq<>t%GVVyU}YFfl|>TzBs$mZz^4#s*5fdHWgoD{Dl
zQ?GBE1=$3x>s+YkrlzIsQXRBtRpulFmVlg^qMB|P`{N!>@#!6{;8P*I;Hog(`AeC1
z?~Za^e8jW+SI2sH@+mig$#&Xp-y8@k3Q8W@U;Ly!4SgHD_)b}(WxT4Y`6}X*Q_B9^
zwHF=h9eD#fsIyD&i`Mlz{9=?e!|vTS@^*;8L*Ey+koITJ8K~2UvxE-Zm{I^Z*%6lN
z_V8z`k&;fWA599a=nuuL^=20qv!twDTbrj&yNGZ8@T{o}7bN5e%3H2T+g^sZYWgM@
zfg|a0dsfHC>Q41)eX$hS!0N1c
z5#g)gL4YzY((u%zLqGu}j(`g_7fm>LS_SGM&|RB|y)y|=^E%qxm0$=0vKwTz+#aix
z*YpHY=l5cq44QWPXMq~B8wi-RhDcvx;B45%a&4a%XN8;lRZ=Bge!t~1}TNRARd1B^^%|^
zXeptk6X-o@q)^XMB@8hk&L!*-zao3GgYOE~xV!a1S4?eOd~1#*D;+oCip#-##>NlN
zp3Lw(BX&NWE;{_akz8XX&m~hGTcMtUpeTPNl;7UZXqTKpkmkrAq-74EJ5X+e7y&8a
zR&0)ZCNNwvSOWObe8IhpPx4=2>kW)`7MwXHe?X(Hb|Poj%MELyPH;`Oy(1lDBcLd>
zdoL(BZ$pBr1~kcV>7D|?AbFLsfIU^W!bDosRV}B
z{BQ83R?~PWZ$_CIPD~7Q&VW_~>&`Fvd~WT@s-hj*b;l+_dzt;DB#AhoZ}qwmxl){)
zpNI3(Xc|T284Aa9Q)53`8)z9Qi{U+r1z%SXtslyuQg|%|vEM(jnOsjw=IAI4U1bU)
z0BZwO1+WZr-j`)a2PK=eF@BLE+>Pt_3hSgiZbX(8>|?-`0s9z4c=VI<2bG>wDg#7x
z3V|FvEZ!TuixP7)YWU%YgtS+W!KT-DP2KE{bTZe>657SbjKOOZ^@OQxoKJl~FJ%jf
z=4dG!;TaY)Q2isybHo(HY+w8o>5PI8rPtyS@l~7VOl5>(@W(wU85{2r??jBjsJhwY
z`GS=#r$Ab4B6~M8D@b5a#M%B>NCRCr9go!V(C|q--Q0jhs4bO0+|HD4s)v{(i7YVp
z+J9_A-TN_=n-ID)Fk?7@qXE5uzfcQ91WOqPR4ex_Iv9DA@AgPSOdzkNI~4|Rucu(*
z7fKc?
zT)P;g!BSmyCyc2P;i$ff`HXh1be>WHwHMmLQ^X$xoY$c%tBkjOS9TnG=id*lVQ4>A
z9kCesp&-`WF!6ef?Gtr*O>xJisjf%;x#8|Un&LqI3fi=WkMp0^5-M-Fmg7m-22Ev^p@)0QzkFt3xmg0600h
zxZibiQ(f6ddGd3_QjR4JBta!Kh$J{rn;mnH!X?0&Q}%w9Ru5iI5SS;E3@E=oyqL0j
zyjtG~6_oo+%Si?>XI$Hw5s|h>8!)SG+dN5G%TUlh04c!3b=s#
z@2E~}MSJq*DIP_39Ydi)wO)#zrOC}F_~RHx36BUI^DT3f-g8NiSY1}${ql)TVfhTo
zU+4yPy`z_pI@}bM$K%7+yWXwFup^o4tC&o*zBja?wt*CZuEDgM{VGoowRY32guhVvJKLO<)!*v)xOy@WZbdA4Kn+XL8*gMYQli_z
z)?M(+DJc0n>@^f*6Ii_?9BxK5Q*REOd_oU@ei#T=(5Z$(!BOV^{a&zD@YDlhM_<*9
z61bNR?96QpxvIVUdv?jgLSG82UYC;|n!I_!Rhqv4)6iF$kU^~vUtT~2C*5T58WBQn
zx5-{M<@jyr8W$t7r^Wnc1oDsgiu$WenVPyp1^(f=tgin2H6zms&59X5I}d|5PIM2-
zHUvbRb46aLN*MqwuhY5FcJ&`8FyPNCD1Cxv_&&%5!7G0qIDmkFxC6yX1Zo2cfNVHG
zZI&cvj>wWs)aG?_@9Yv9a9He&PI*sB#REo>I1LnRdxGaCPYLIhGqwXxy)Nh}C4374
z=DgRc{5?Jj!~gI5&)ewq$beZ6oboNubkv{ThBb#EoxRrHifGw31NrSJ6enG6EC7Ju
zRZ@iyjP)|?GfWb)E{q~#as4#%|IxwWs2Rs|i9Ki6J9=NOl7W-yo~l&nde^;^Pouh3
zhK`bu82F_?DZ?N2kz{LUC%Jc_RqPTvNSXP64~PdkX2PYLXI(A)=oB*JT*zSw2U8X$
z^k-HGYwKw1IfvR_G)D6YVy`QSe2;WzY0l)_!Xd^?NRKr{_-uRU_+$WMVh6?C)pFv~
z4;evD(tNc2E~ovt1oa+U!jqThHw|VYKH&7)PO862?P6ICnvS&{OpYZolXm;`5{+Cd
zlhbU3oku_f1y8&MB%9!CyU5A;9Qyo-5##`dg7mrv(~G>vV|wh5a`R%qE}r)BR_A(1
z=(l8+aBv5`rz8&fR1##1f+R&VWpWR?j}2DBDC(gIiawa7D5RemG0g*QbT4SWXfp4TyHnwUrj@eZVry
zrwEZM{4O|Wd;GBiT<2#=Zow&w*aU!{i#w3CiItW;p}NAcXJD#3XXJ%E7Y^~fg!=&K
zL&$=$u}BjU5siT{3wJga6)br)oq-6jlc3e%F&zNEGDSLxg2Qcg27`mUaq-1vrk~ok
zdfT?Q64oV`PCvF)x&H$07Yc5YS%<{)|Lv?CI?Xpn@v8LA&uJp(!>Se>eh{i70n-5U
zEEcb|e$EO?=0Si0xKpu-fJ#LQrB3bFNfpL_G`n(5gc#VimZ=U};Md%^W(T9dW
zd*o|?MIa4iU6S&?=mNK+ULc2{`^(?QE^}{0E%e!6(O1k#Rws0)!UmSsvX7M(dtNIa
zOZ3BFU*g?&cFE<1I2t9BOAx1iuW30WS&=*@W%34CwLrRXhLKU<9%MR0o)lc_*JqF3
ze83yf6a2#ARl~3eS3JIp!?%MPv_g7s;aRqkG@rO4^RsX%ucKo$O<90@PDUa_ut+Fw
zjpEpn{W_+!a9HtYxSo)b80TV^kVqwR(?`cbklzEq{3M9*@*D?2p@8Q)-7>UEo%)s3OCVT>g3C}-3Fl_;OGg9!-EHofAoVb?ZAVT%zpKGo
zLpO5F(JkPLimMNbUFg9DYuR6^f2zrY-fxU`77y#UW@hn8`1z?W^_c$TYFB97Jk4NK
zc{-0zC9l!GwFTSAmCe)vv0GbZ+{$OHfr*2%u+RtLX(aO5Xh$gGzt`4+
zQxbF*9T{j&@TlAdy1w;L@z;%FXdJ_T20*Q0uaK!u@KN74`Su(RKMa{*;o<02j@}I{Pf9L^ZKn(RW4`Y-_2T8ZVK(twt1qU;ZS0t
zYSFD!-l`)hqhnbjxDmlNqmF-CTSvfDn&5YqB{rrXQ&K<5UJ*jp4F+R9`2_mcpo~HY
zdoD|5Hr-jPA<7@d*(8c25gaE>$l~}RQo9IuA#1Q1N@`Ud&}88l;vD31gab&f!lI!0
zf|abm8r*BrdRQqSc(Z-(IKhK?yG&R6$&)JzfrY<&OIOtAw#yEQ&(C($1W=Wph-(U5
z#HsSil%6o`X=19MEjVZ1?8^FilcomlIRgA-=64q*!f>ex>iFSq6bAie*`Ze`uxM;F
zO-Y3F9DHkht_tT2m~k?-?o;dZT~zNvJYUGN+a^zBP(wk3S$Th>ijHk4;QrN9KAn@R
zHo3Y3bMIDD23!=552F4+y~a_XO$IfPkC?na+b%w=CI*zA^;cagcN64)a{OuXuFZOAbipfy&aChZo=PwD=h}CFAAR
z1P<$p^k)ppLV5Ff6Mz3`twX;h@rU#QRms&PA@^S%Gg1PF=Emgm&U*kpMoQ_8)!g~*
zY8D*aJDo?upc2y?*YBE}HN-^#!hwN=A0|@U1;iO=B+ValK4RnsW(4a*YH$3$mvy5i
zLASS;@hE^~_+V7yk9u9%b;yI(ki`sli3e`K7bt;MKQ`zLck@%FESw?dpzBVG4Ns6-=m`n1xYgFgIQ
zwC88@7k-!&(HUA?dqE)<75#=#Kxbo{LVni=j6*^nNy8-9Ct
z@Y*0HjU9O)D`ZkaAi4o-5G;$0|7_Jtw
z*!pb%gvcy7MZO2(HCF`DwPi?fZreeGqW+*iT3!GGv
z>>0SGHa-7#eDOx2dY*ZxJ+<7}(>vmn`RHJotvbE(cu&pEq{Qo*u-)sq)x)mNEW1S&
zY-a!2B=@i`%ZX{u?Ej{vX;;rfeg-cv=WLi_eM$t_+Mm;sQ%}*d|cgEl*FxQHX7ogn4F5!F;^q4b4tXs>&bJRkNR<;;k8~-5i1f6wX8D)M9gdykOw=}A@yNblxqt16y%#Zb
z$dbdSdZA^PwN=~}9!ltDktq7$`c6!tA!T-rM^bvH`YMYjBUHVTHfeI#?s{n1UA9)A
z=^Z~!N%JTXDBMXt{J96;KCzH@zm~arG6AFlNQd)BNv8#e!(bk4({GP>pUj`S=4U8Z
z_xo?RR}L%zK{nWK@R*$ub9a9}mzU2}Q_`^g3=&0=tJgah$B4q+$5+q6Z$M82*K5g9
zIjn*wY74jM)w3_3%b!)gu736Q&
zgko*u22x^dS{ZL8kM%8m6Kt9O(ISV_m$Ph}&V*|p)Z5@jk-9a@hGdH%Z+_z9Vq~i}
z#?|C$aV;w=(%Vj<(VZk$>+r3NnEA0|L`G+iNibJ(6L$MQE_$8EM}4jj5tkjnzlYp&GPvL)S|hDQ{IRtAt_8VrwxYUx(+Hy4cR8xLN4Yk&R9pqj<6(
zjFgn!7-s3!GE5lOH`28CEt~&he#a?fp2Dq4UG*1HZEooqZC4j~wVwpNPhfU_ATdNf
zgsBz?i6OSRB`*OpecT^tDhRg#2!X1gOUxnXQE__|HHDzmyEqRwDsDQ!c~yfM7%kGL
zS3|eJ9?K}azjpJ~#|b{u3wJCx6qebE9eYkKcd)$h6koaK4=EWGVtpeg%h#^iJRG+l
zg9c3oFE-^Wy|+8xZ@Ao!HuqqpYTwGXQKuPfxeX0Qcl$}acz
zn;^F$wDSQn{m`oIJTeWgSjG)u!+|0G#^K&7jww-Gdf5o+|KIB3hNhI2(3HzFualoU
z4uXOEdbLGE4~4aVpjXI?^=2&jhB-lMa*g(o
z+DFdK+4=8DP;TbBnj#xG3NEE!SrObc0^$hbdLkksnqmFG)BFi~RfsC5wv5+{^Q(u9
z(&9(UznPzGy{xM`cIeRZ{BMPcR`LFdl40U8E|x=
z$^)Xv`(W?7Y*6qD@H@`4Z5&Hrq(BNH9$u6G-q~YWb1OCF%*SUpa;BYapm-$4zgnKm
zGca2s{duq4#nOogY@oegxmbHuPn5111i$+W!$W@TmZ-BKFZNp&ys6q~j|W^4`IR8UV(AtXj%OrHm-
zGx+};L7LnRcVg*i7kFOdw;rB!<0{?Y`z4{rDw;i)OWH@*nLXtGU0<@$P_>7vmm(mU
zKU&Hp2FOq7YwR~Xnp^K`0lF!zL^6M2?cvYa5U981YZc@!EGs3c`mNz
zHl$y!BEv2N66@;=lLP}F`~(vEin9vv+tHB1NqepR2A&3A5dnNkb&ZV)4n}8z$2krn
z=M-=k2-^*O0Bu=UlUsQR$ZQZrNtTci#M8G*KgU19S^Vi#Jls?ypDPcx($YOm76V3opG&9E
z%iB@bkt*w`}^<2G9
z-V^HU2sLm1uYuvTan|1cA{W6@%he0I;;(>E32Ify&u?oMeg5ng`ZFU-WrxOvN2ZdV
z3=xMNyP}tpzrm#%2_EyGH!Q#PBRqn;#g6xKNOcFmYw#EuoCQVpdFt1L$
zD>E-k(b38ou>0h@&n_S&)`+|0T1^nHT&BA4Qg>)nabeG-=7%Rd?dCX69&B4>Q88Y-
zBV?GxP-aZ@Wqm}y64Y08*5A2^DYTD`=5nKe>#DQ6^%DZQykTyOuJZdEu%|~-6gg+W
zg9mne;HRE+m1KpzGyz#BHvq3-Y$&YNau@{wUfg#uH6jXje+QI%V2AmxdHLX4Q?vMh
z&46Q@*yiEm<6GT~b}d$+CtXWb4aGTNXZmI)LC@j&bnLT}*#T&==3r(FqT->&~
zJK@~O!}V}o0cwq^(oZ?BkkFASX+8?^)m?~nB|3-xSp0l%@uVwP+>Gr{Yu<|!y&46L
zP^P3@Sw*{wFhynNU$3HpYG~he4YMFAj~U-F*2%*T(19jxqA@s?=I|aKi9>K`^}_^T
zK%`6owO|J7{OJ&GF9THk)vH%A*nBN`M#vmageQrRm_Z;a1QrHxJcUVb+{LxRD~1Wc
z4mk~X4MRbzKDJ6$LWa7Se(_uHc-+?$Q9qZGti=B}bBm3=jN0`4X5GDG3)S!HZWH(V
zydmOwwQ#H9FI^(#@K7A}Elo`p(G7>efcZ0yJKNcGbmxUqP`E!svZ_=9c^4f!?_(9&
zsZ*zK8T6a0DexW0scR8umdYMOEgqa&)c7xZ(lbIa8&CPq&4Xsx#E{1z-$Mt=W>1xI)5Z)fD9w|ZV
zA6$!bNq{H-9T?Z;5mtW$mNrp1Xm1WGY-q_KejX?a**XtU5=N`o{b6wvw}Gt_QS-ux
zO(^ON9qDG*r=NG@pmGXN=hioimOtwPhsxXJgZ4k3xBbx^D3Y#X?@6}DF0fH9|IZZQ
zcy=zoNuroZ*7DQ~IdgU?vJ&1zu@2WZ9@omJ7#Y2R(J^u7PU}-=12WNus6~BlWxY7a
z&+W>LLz8tLt`$qyhcG3rQk~CzmhW!+CH|MOtXbjosrz5FP^6*@|4nR&97@DRjl6%>
z()^l7M&ZBRAn#qB7ogf>Y!oG-bs8ph*7FVK+@%vpekLT6TQSPH((9-C_n_){#n|z+
zo4Ga79ixGYs)BQjxitLaPgu4R8ofE^!}d%Ez{{{E1Ln}I4qwQVm&=xAE}v6d!#j`<>&j?
zi|2t{kc_l#uwhdHpMT~_HfLb!SbCX2%>hRZBr!F4^<3RDN(Z^!xl4qI?wS$hgV3xpVsx)oGWbT_#_&7!_vw0&G
z5W0l&=X?kr9KqkiY!{oRZK&3L6nDCxW6~Ill+y=U5iK+5l*3@)N1Q1wR5$lF4k)vN
zG``_KCdcUR8bPBgx(Qmv$Ot4)F6Hr(Nc3l*-d%m!cu!)CNklK$0-7r`9_n#yRXxK^
zgB$$F3C<7FmZ^6*i{mA+MiDQPd>m>twQTGm!6QiEqR!K^kEZ*;B|p4sLF1WVg|0U~
zrl9KAy3OhU)PIw*w3_h%YRXo5Eg`3-_REt@=N&R38hE+Lphx5n!l{AHYP8V%S9jn?
zi8fLgzh(3>n5#LFQFWS!Vl(3)L&87lLC)$kX}@=Sf3H8tIbZ=@Hjm_!829WrGu7fK
zZ4#vS(T?#u8-0DdsDj1+840r{W=3KYG90)rwf>c$d0#f>?I`)j)|;#*Dy(fo$ZDZV
zPx_`i+?Ho7K!q4S32Mv(4v^07oWfvm1O;zhlHY^b-`oiugYb1e=t0pGL?9o*AlF$|
zWV_T2#mj
zuNyd5LLUXlOt-Du(%AJ4v8ZXR^5xHyVsUks`D0bn8gf}8qZti
zRyW4nbkgBQ*2$ERuF}gV?d%(Qu~ZV2tiA$e=r10uFLrzNVz5#<3Rr^`K9G8>9dHCt
z7spkr=4Lsrjobw~;*aKAxIxX6mJY&}5^e$GoKi$TMpxFY)}Uj~v^PtInHB
zC3RiW7>P`18eZdkD$iEuYQvt@@(e~aCR}2@O8UYaX`Z>LAPN%kH*nVmXPdz4V`$VN
zcy1s9mkd<`s3JG+M}XqLl{OoklsjqE9s7sScUX?cyf2v-LGO@|xugu(>MV>-Q&Mi^
zru?oZM}`O5m%%7L0Y7@fhU4H*7iK$@^g8^R&OX2Q3@&y$v1EqoOlex22WOiG-Any6
zp|&^`zB+zUkK_i_57&I3lA4MJlnclr`X^r>rF)XVN(5>M(AmJU3sbK8^3|bkZX7Wh
zc;cN1C!^9$(aJlLk0t3i0_d~;6}=v&4no(adl;X3Y2S6DCsGR8O&36FuC$z_8a>B4xGy7>2L5e?tx!rc=d_m)njWn9`4Cj9re
zeRwiwo7j9>NRH#nDyq8sZSI!U`gFPJE1mA9CXK@o&MWT|RVmK+wn)!p--pd?WRu#%
z&(Hq`8Dr9}#O?&-=>hRnp87{u|Fz(X9p}_414194v^C<^2-JEA$$5&?mD5sQHW9sF
zawtI=Sb4{IS$^6f!AyXcgW}6Y_6z$_G(^XkA9Kt6#W24XB`=dIOz_#cH+HZt{${nE
zLct7vdH-k6W1ANzmb#WYi*q)NG&ebN?RI;J&4QK_xz?6my3+4D+w^EFIZi$VaLYq7
zkcHM?GG2T^o7$z~%*Mp@6UY_N*EjF`662bEMmgcgCQW4EBx=$wvA}VGj~bf@n7P8Pv}w
zR}rb-+S9QC^U=_K730DE8o6%f4&zrrD${?NIhj&csj8hJqw&9Yje2^=2bB+2cH5=h49lyoVLs77_4vJM}8ubDPbGs8#Qt!
zm$D{mjXWFljpU%}Hfgupjr%mjs1q16K9M`)6YJK%z@h;2J&w;l;YtiId#HfRo#7Fz
zj*QSc@eZ5;%Aj-uhZSE!o7IPgb2vn+6oFV;!0gvZ*B_8^Dy(g-ee;#yZ`NLerS4A=
zIK69)`fidx2B8K>QbtW|6f!xKG6Q8TqSV-$h=hgQ&A%S&R8~W{#nl52$OALf7fa^4
zLn}C-vx7|Q^sp--KszpZ;PUL=v#SavS!%!xdOIRk8mb{-H22C|LEUGasqt|0ZnQ5c
zsB!Zk{Dapk>q#wm<6;^v@Fls~dfau&Qm0Yk44^%19q|cN!dTF-9FEM*RZtw0UK^3R
zaK9!$2qxPv%@64Iy15Z{40}hs>w{Q7Ti_QT1OcmnW-2
z1(1A^AdEz~!kzKftHL&h{q`-g@oT+f8GP0~Cn_yw@=&DI*p5#XQJJdOn6%Fex(MLP
z^u}#6Znn=0HEf9IE-c8Ke~FqJvr0W5mnZUI>6ZDQ?;qM`KAALn-F95c&T9xSW3RF&
zDG$>RJ%$ha#KRQ)Rr>4XCo=)H&D<>ZRf89dZN+!Z1;PcDZ(-F3=&*)4x-ekjWzz%6
zX(wRgVK6?|gq%L!^bi*>lFUARq$L^!&k2;BnK-_b
za(>g{5BZE>ETgI~Q`-%3imMIY&sI-9d#UEIZqT#$Jg~|>8+#={T)Kc)yb4fy2!d!(
z_&2o5d~cmO-C|8gMV`Y3l!$I?+3wCGejE2Enve$F5ABcEOYKJ`9VLVs)@=%tseg}p
zUv%@%yd
z>WevlGuin^unO}eOWDQEL34c#1$X)uDXL;n1lX1d}$d0B!fGqI2i=nlk
zU?5e?sS^^nUg1CA@y3Uxdu#)5XHc5a$yL!h{YY1J#<2iZ;(MrwA|_X}`g`M$P5!3K
zW9ga48?oMdBZMv{YJ0ovn5A4E({BWNEETR%I&|Jyt8h&|Xl9?i-c>
zJNDV?#GJ994_<@1wbJFY4!~!jy(2YyE2c=>r3-^tR4#_u{k;*wo7i+Pw?R6SgL$*b
z)jhSA8xt5>MJuo7p;Pu^UO<+;S6on4TETRmT_o0T6d*=&=@)J
z8B$ogff=Nsn0ank9x5nI9>PnDg)dUlh<%@u#$stq#>1p$yhCL6zRR6MompntMNmDI
zlmEyHbc^MR&xzei<_=L3zJ}uXS?1@Pk&g|s<~UEc?NH!$2XbFB8TCvoFHYp^f0-mS
zYP}835vK5dcq+71rR!X2U)Pe+%ZeN9A*QrR6M`39YZFiz{iU=~xTHJ1-}KJCJN#QE
z#6Ive`MTMoJ-J$J_YLtPkwWzKhKoeBZ*NdLT3!@i_*@Y&bVr%QuskxECs42TL}!@9
z71gKSA|tfJEzDR|v^a-4yWIeMW9j$09Y_WHSi7%M8tzCT(zQH_V)5x7;Ob!&`)r?hFCao}EA97g{U(4Xs)pUiUuH`{GW_
z^9DLOxly56&hf8N^qSx7UKbT@ea+SXGxxptmz&k|+ShTbDEy8}P^H&E8Rf|27GU^k
zJ6TaxZ1S2XFJW<0I(Az%(ybYSv&3O0Z}AMhXI!RpbvNU@wFW}|RtVLV@K5jN^I+${
zXs@r@GOUfYoV`^Ll8yB^9UnwEr+hRl^0w6DXD;0q4G(j__iHW~AGa3}?`2TC
z9p`iXKspe&yE788s~ZINY?a|kw@JS}hn10vz!LAg|QRl2!|F^s>!R(O;7Z1nWcE<8(@_3z{$gBeP%p8TW?2!83YR4-o
zJ`)N%Sd3(L_S>)}%&G&I_Gm6U^n3>T+6vBlY7g|rV_zHDY&W)mT%`|#rr?^qx1Q(6
zB`&@daEWI~Exy2AAE_+DcCd0KFL-)Se1XqN=d3w@3sLtAj#B2GKU?p8eT=PyduoHt
z3$?W|^i!&`pDdbhZIT4N^sh`g&Ln!KYU#P=hwlq2us6=~%cfflo}PEVlFs47pEb<-
z>7_;)rGj_yu;FiiWu4puN+ma5oK)1QYxCF?Bhz~L&i=3F^t=c2tb=_SCcMpCWc?*<
zX?7YST|$4}*NJ6R_^iYp?YQODG?Ag9e^C&
zJUl$VVPe2E!bT}w$elPmSU$|ieq_?E#Uh)+onyRl6atdt$}V>$A1V&~fo9g-2594}G!IeuVe4oM)Uwx)KKj
z7$&?^zB!AVolswV2Z9nQPj*t_GicC>&qPlf8s3HcYh>1#$#OMUG?Zf
zHvW8-4iw22j<0eU-(*Y>#47
z^P6HWJ#0V#uI{-T-12R7NIpeueAHI__lQQW5E;v=h%hguUYj#o+_tie0n3}4zWJQe
z{LzGj3?Jo;`1tcPpGjwK?JW(Q#rgp)4e8hbu09qiK>%_(9{^!6=aMOeJPscxFbV)_
z(1}iW9zR5T8fHpx{qI+suuj+gwyNhS(WZ3K?A83ruF>u*%r2+Id=}5g-K*2t(va?b
zjoOc`CVC%g`$*w4_oH#Hi$=S`4Sh=jY$@%G$JjhYQ4m{;9tTS_vtnTaWBf~dwZ%8d
z-S95!OyX%+Oe7rUVmQn#QmBH?KjjAP>)Z_GSqjqWzP^0WG3
z=4MkdBtzDEj)(>lSbXK9%prC!UY}Ne`x?qSC=&;{IHGk^L_tshbP6vL>%G
zRgn@FI5wm2scWnrROLx-=yZwc&}=-`Tz?tg&Ju!HtyNC%F-s6tmfAeQe;-1
zVa3z0bUZXTDn@l}WW#xrXoi;_2Ub6|+f7Wux!au6aExzj-2Iq-zw8Qf)XlfpXyV8<
zjYl@+kNw@X(sPUl1{pH!5D-uw!P_7<z!IaGJ7i-Mb>Sr}M6}v5ERFAZV2>-<=
zJ5DcFjO2C?>yeBO3x4^Dt=n?>p8Ju`6YEjz4jNklTMsmmsx6tBVd1eAo18D
zdWnYeqB#BdW_PZx4G8#=8O|{=Olp_ji$x&Q8Bj1+6hQR?$W#bu56*sTI(VmSfK0#O
zQ2!t(PE1FYAxOLvV*Ax`pb5zm^|fDCB>TtYyg|pg?|D%}#9kYFFmlmH
ze4ekJtn>}ub0c=Q=*#F3M5HtIX0;DG-ZzGfaBN+T97%czQvjAHZmW8L2FQ
z5gdm`zQy`{B)@6UnHhbFq8~qxEYRw5i(H$h{0N;4=H7cty&LCV*k<>!2?2iD=3x19
zVwDuGLdH!R6*;8iDA`7?lzifi)O1OFm%9#qHF`6>NJ+K37lWxED0a3FTJBtAfdhLr
z4JVcx&djvg8++YDR6^ZtAoK42*>wt0+^)f`SCKgWoP7X8O#z^p2I(awzY)B^#|Lv+
z`H+$80Q|Fz_e{&1*$$ksOPziz{cpp`8vmzb?o_DlNZ_RG*xqEi`8zQ){Q=ukc+r0s
z8t*!5Id$op;Ri^HKNl1g1})EdER799rjbaF=hQkt-dK4lieG3(#Juy
zCBJZeD$ZByH*I=xvDw(N>OQyg>Uin{FNwXkh~#OzNYi-L9p(LCM$K);Ega)>*K-
zPRS8tqvr^u+Z&a@xdp3L>b{OmR174+V2yrEgA
zS-QC((>~7VQG;%kFrv%#*~J^X+nO7Sx4l%9e<&EF8)(`g$J9~S_O^sMszq(Pg9hc%
zBQL)rnpGrt;lU!A)&8>qH)D?5+5}2Gj13ZvZ81z8HXo~f0h6n@AKXvQ{N6qAzYrJK
zo5^##H{5^H>BE)%AIu*YQc7X0OVK5(*^~1$_FQJww-Z`b7wG?!Tj`sSY=lScU)r@E
zaHnLiSa$1Ac8skWk(Juh5x7FpByrfCSl{i@tgx0T#*$Khyxm3S@%C_}qE971vp@gU
zm!>GT-t_=l0b4ysUzmA+9ma|vFjJT}GBq`Y%u9g|9Re0NZACkZ=L}!4KC<%Da;wp^
zJk3f00=PpzQtL>q6(BDU>t3R4}3SaKRrrFza4Z&Werf;{3D*pRa~0wp+cvF=L@A
z+?8&h$l4~UFR1tnVo=t_2+?Z=PBM*x^`Q3}bs1H|tUHBD$1c+piH*X}(?_mbqU
zV;Fq?wta?JvvAAUXd}!`Pj_+wk0d0lbZSuE*4kY-`6{9!d`LD{|~?8o#i+u>}>{
z!Une*O;5_Lr){_7ZW;=Hj(!=;nvwtGVjUVeJc?{=Y>-*XFxdbZE((KVw(s9{BCXi?
zlanuQ_i-&vTb%i*9)s7Qr&Amf(MYr^NLMg$@0)DSP#yM9DGkWNnR@eQ8bVbg?vQdL
z=+@f1*RCZcbD~`-Iarc_V9CI@=$V4Y71cy%GL81nhR&b1@iL_1S(MMUtW{)r
z7QXNK@nG4axph!AtT(WuU0v>
zUj$~{{mV$kF8it?nHJ@(&$jV1`O-}PX9GyeLXx{rNgiC-To>JgB5BcM?6;(sr9KO&
zjFAGA>OqZYhy#C98k9$V`gv9jN6boPg|nMnnCXuF8ZnbFpIPt&ZLBgj6Q;dYe6$)Qz))0$PZJ5XxHx=M$H$_I87bqu%e)ET`M-dnA?~V-aUG
zf8yYO$-4v7{URlc)XH+
zR0cgXrwQW(H^H+=#|u_C`gXhv$#vHSd2
z#Fw_+ZbOFCR?KAA)b9~w*?Dd7R82vSc$qQGt$FiPWc7#NVFKNgmGgqJ^-ABQd~m$t
zqeXcHSz)G-)(M~XAKG;4UI<77DuXF-n>&4dCLtiW>}_4F$;GRlEU1Humn3FQBMVd?
z%6SPKXdWiJvRry$S8_;sSuk)n%eG#@ia5w_>KBL4_hbac7i`M_wITdt}MeosKFRDsEA!BZEqnoL|pcdQ}xh
zK42X!;gXp2`=i0sQIur0J`x;k?!&NZI927V)?S{TY2>kZ@=fQR*eo0}E&Q@+C5xZm
zp*fqLo}Q9{K@3!7W3D&-Z!=kOE1>@@i#&sxhV_p=b-4s|;9$m!hWNI=8cqu|Zj54a
zY8z@jx6P`Y#0Azz&P!SKMUR+s;!BBKHrO83N5c|*`sH<$Frz94Grv*NdSX_VP5b_}
zG{OBZlWAHNoqF
zEe#O#>VfL%@+s{stFpkXGs%_O{M3=a_Zjqlocxkkvyntu=Ha7cT?
z?{`b?mKfVACyG9Ay%fnuvk)oYv=hw|%fGhD<})qXSg<(TY^M+iHJOyzkgZMERx#H-
zq`+25XDOevc^Bh`PAGh_^V|@?Q2{339iS?GVKtgRW8ly8)`qoD1+k}h;c`m~8DS`H
zdFV_qgjSR|-%b;WBLR0xFwQ~zfM{Ggd6;H9*gYj04oaZqNhBPB>0$Xi>VSy+1P};
zj?}c*=8dw%u2k!WD}1lX=@LyVprDC;n#^-2l#1o*>(YwPLY9k(kJ<<8jMqPIek|qG
z8V~_Sk*{>%y;pOgIpXhO861lfr*PSE=hO8K%@HCDhuwlXlHGvn-CY|AyLZHnII}}f>*f=0Aek!
zvsc!ICz#Ki{dgWzsbAAqex?ODeDjB>1mZj@luu=rB18Da23hZ_jL8?YxC#WGW#0T;
z=Hna9tGbqK>BN*{-Ftt=hUQ77#<$k1xzb?A5pC9JDW6xVUlP5h(KW!v6gsgbv2}B!lrJsjTPP@OeI!PDU_Pl
zIKY0j<|>IM@$hQsvCO<4kAxr-J7ELc^j(7}>K!tONQjAvRk<`Tb14y#3NKT>R{8@?
z`eE&?yv-_{t|Bq^z_ng_ze7TnZd|9lW;W|T4KzpE399LViiZiBZWR6fnfuL&=xi#e
zEY=D!sq0d6zzX$VDOqpT13+W(8mSU)&2jM{%z
zOLKyRVRESp8o7vNRFZmNx}0=eaOO%3A&O#w+xA#@mmAqDW5jX8Z7lfBE6`U9NUcrU
zwOgI9mW-%nxOS?O*Fo^F`vDJG?E}uPQgvA(Y*=B|9Dgni2m!73tjVR+7TQESA6yTY
zicn-#d)9wVGqv!>N31ukXMXOf-X)vKjqhM!C$?L=S5V_Vy%WR9;k}SX=X-97Z$rl`
z9%sm3&)TWj$|vwaOXG89%TTjHp){DEwQf=-75MbaZTZO`3EGs;P+?PvT=u6>y(j!zI%
z#L`V1KoagkR}i|
z+2hqH&b|t%2o~d=ZwqB&*)p?&LmYQ)s@k(F7vONqS{NOJ?lj%|A)r%f{{y7pHDMk&
zrv#qJwyXYx!bD)qWlxu(j9O-{Ts`ruf{bbEcByyt3iMzxz)4O
zb`B+LxJ%utp&b+-pb8oqlEY}UK$BJeq(qbMGFwX<@jw2opNRm7HSOPYb9n=?7^-Y&
zbv!T6tEVXhyUHf0dib0|A$(9z&nalTt!nbjV$!zV{;u3l3<8cTX@qYe!uSfuR2W5n|z52acQFnkDU|
z+pE0ixAC1rh<XIXeW&|UJ7*=UR>M5w-{Jv{;p7wrY0q;#;qz2m=Cv%0SqT);CTID|9z=0oMVkN;7
zKD-vYxkXnfxWgD-&7gPSyPv`AdV(|FW2dXZphxNl(3
zncMR&iBidU>ncMRfOBNp4p02+?Db$f1$;5;|
zYV;kP+3fkf!tUMo50F4~{|7+n-sK2XXG)DLYW7piyIm%b*38c=|6HhFY^DHMDwXRa
z-AD8`0`uf?Kg$Y{Qf*hCtkk=rlF#$)(HmtA>1}dDEZ!lc>J?G__*@yBTA=Mpb3od)
zS}faeAT~goGdWlOvVo)ECTL-10xwZr9T{}_X{$RV=G6JcG>6UD!H(I#qe!lTig6DG
zu1i0Gs0bZdlqpuwfw-c0e&h9caui4glmwlO%Q}aD{Q|URR7=vcbumW(^!Xe{(XhpS}
zwzL@oZ#u*AiUEYX08oX+YEA+EmxhXh)U2u)%D%-rV2rG|BPZhce(l>dkZlw&sN^S^_r#R`m$8Cz~?arWN^+vd<EgRI<^-MyK;9^<PQdsLt;IdwTkdu?c`YQl3hmlTs9Fo`njGtDh
zPe6aOQoL_jrB8E*l9t#nF^V+I$Tlt6H*MdN=CtPeFHLi+=)SwQVFFAq6^nISQm9kY
zyT`OcRsBPHC6TePR9X4X^SIjMPNmQAg6q|xYTGDcjwhdjHz
zZek~BZhh0JMs_sA)4!vE0M@Ivei4TW7^1*7{k7WnXo>dhBD|$K5C~d0H6$jJL|rwR
zBM(T9l!yMO`sQ?n_g04_uYaVkqh&jy0cLLTeRcudnY5GnZ
z8%x{w(m*|>UafxE-t1-~m<`B?$iF};#w9+ij5v}Yn9#Gb=hVR2jg%{C0O1wDy@x8-Ucd#^fkY6d4+}yzSp{W{{ZbggaQp1RVc}OOLJkNWx{;hs?V|z=W-E`^F69
z`I(VcWRBV0Bg^MV3x}9(Uh9$sm2?^IlbwF?46VS0?_da8STw%3%X$usFj)3Y{EkF1
zyOQadf}+4@4o)kyg)O;p`6e%TDuT5B8}$V3DyA{wMDoW2jdDoqR!d?S#m3)%UNMj(?#L@|BC1i3Fj?oF=m
zJs{%rPbC{pI9{`45FG+uR^p?S73_OrDUpWr1lxu#U+CNkt)IjX<6-Q@V3OE826009
zn9|3i8~qT*@_p9`^Q^V}qVMJhbr8RZyGOw7vq%O}C5@X~DbC!GS?Zl|j}>(g3`3f6
zyHv~~7R?jbEXne9ybp!lwVHK~hMwIym2@lU`EqP{!rhtP;&!_2014onim(
zV~Qno{x!`csBTks8`mG$N)!1{)y2}8?qjiEFn7n)2ZHoK5O_OG_~N23q>VQ8c23V4
zcqZlHqw|p~=>4H()K{w>8t81m6~=k0ZOH4h^d13lu;?R}OQ860^cpb#
zaGL5^mc1QxW_Fp%QT-dP0PbM?QGOy%u6X^+3^L*hisXQed$L5X
z=4Ry(XraVYdEQ6AJ!KFkxZu6&-vTkE&t;f6=--wZcq^i~A51A?S%UZi9pV}&DG`$W
z@+kQWCwG+6Uz6Jp%Zmu)HUU%(f^bY2$X~z8ZIYx4LR_Gjy5jsi^_B1W)%+bSr?As#
zuqzqXo)KoKlLqqhR(72MbCF+^g{x+22FgoE7yd-c`GYSKrW^=ef<8*|#PYJTuguNO
zQH@cxnKm;1Vry7Sk!@wRu)ISYaD!x7vs20>_#W|lT*7$y4gd*JiPbzmxF5#0@w~je
z`uVTSQ>5Rq8jiSLCjN~Bg8Sh(3mrvx`H5eThQ#{&T2qWp>_m7#lt|*X%>;`b8i_7J
za`W;4fA9e?r|n)Rk$n6UnTHMC7wEi(?N=IKxj}N2%!$2p{PLV#;gE?MS}z=E=@thF
zbsorw6D2vV&QR0*4wM3{SpZR9$m-e+gJ*P@Gbh4qbw(fXgz{c(mnkSIE!o!8M~;t<
zF51>q18Np`_E$q+WKlpP-#=^h9+po4=v$-!&m11VJ0~DjKqKPB@&+_;wtyP`!rmv1
zI;y`6s^Rl@b^SOsf!Rk2YQixEt{EY)JXw&t!UF-6@;abJC?Id1bz;Gh!Ykj6whPUG
z)yw|gllSSG9{oFxJknW@td(uJCM0Y=tzI%T#o-y|?J
zG(;8BN@!Kz0{Bome|~yivzeRc&L?JrkgoK<>2G}3Y~iYQi;pB=1At1iAdJYgS+m4r
z<5wmq>r2vVaJ?}2yu>I|;TYKSVR%&}Zzy>Q^f~`MCs>Sb$_|4ObU_1r96f;AF*`30
zzFDutdPww%4d_|~qir-RYFH}^M&XcALPfvpBMLm`FU&bsIqg7*jJdIAmmGhz?*ThOtmJCKL$DetCYH
z1WSqhI}>8Ca)6X(X~u4VoR^mf@ZJIb#r?Ch&JBCea;>2gko*DeGk12?#kIEN59)*FbDy8h|D^gGREA9tqnIY4BEfP}>1<_}Nz|UDbU{28~MaG)EVmqMo-rfBE$v*IJbTXoUJw|%`r;^=9M$s79)4-!IL1wY$5!O^0pRaM>U@F){WN*&qievx1Sp^MI
zIIoi`H#fI{xE$<|RYLi5TpS$h{e@ScEE(*12IxhwldvJH;kHu&D50I3s?-Vq*XH%j
z4Xo2GXyew&HV3%qz;t1!-#(eL+nhZLP&)!v@r0K%4E-wYF;{CW^tzlJm`$5bvENMk
z4(5c14u1Z>CjD<@oj_Sdq)`F->GjyoiubE9LUVAY^jHZ1j=Q&qDX6N7Z)j);2_s~woEy=z#;^uWAs*AI`AKIj12)JLRcjJl+|iY*
zg$13k*Utf~B|m;s8TSlS