---
base_model: ai-forever/ruGPT-3.5-13B
library_name: peft
license: mit
datasets:
- evilfreelancer/ru-chain-of-thought-sharegpt
language:
- ru
tags:
- impruver
- russian
- cot
- chain of thought
- lora
pipeline_tag: text-generation
---

# ruGPT-3.5-13B / chain of thought

LoRA адаптер для ruGPT3.5-13B обученный на датасете [evilfreelancer/ru-chain-of-thought-sharegpt](https://huggingface.co/datasets/evilfreelancer/ru-chain-of-thought-sharegpt)
данный датасет представляет из себя перевод на русский
датасета [isaiahbjork/chain-of-thought-sharegpt](https://huggingface.co/datasets/isaiahbjork/chain-of-thought-sharegpt) при
помощи модели [utrobinmv/t5_translate_en_ru_zh_small_1024](https://huggingface.co/utrobinmv/t5_translate_en_ru_zh_small_1024)
прикладываю скрипт [перевода](https://gist.github.com/EvilFreelancer/230fb48329889506cf88c03b8893e4b9) на Gist.

Конфигурация: https://github.com/EvilFreelancer/impruver/blob/main/configs/ruGPT35_13B_cot_lora.yml

Адаптер обучался на 1x RTX 4090, для этого потребовалось примерно 20Gb VRAM и заняло 19m.

```yaml
output_dir: ./models/ruGPT35_13B_lora_cot
train_path: ./train.ruGPT35_13B_cot.jsonl
val_path: ./val.ruGPT35_13B_cot.jsonl

datasets:
  - name: evilfreelancer/ru-chain-of-thought-sharegpt
    converter: impruver.conversations_to_messages

model:
  class: transformers.AutoModelForCausalLM
  name: ai-forever/ruGPT-3.5-13B
  load_in_4bit: true
  load_in_8bit: false
  dtype: bf16

lora:
  r: 16
  lora_alpha: 16
  lora_dropout: 0.05
  bias: none
  target_modules: [ c_attn ]
  task_type: CAUSAL_LM

tokenizer:
  class: transformers.AutoTokenizer
  name: ai-forever/ruGPT-3.5-13B
  max_tokens_count: 1200

trainer:
  eval_strategy: steps
  save_strategy: steps
  eval_steps: 100
  save_steps: 100
  per_device_train_batch_size: 1
  per_device_eval_batch_size: 1
  gradient_accumulation_steps: 5
  logging_steps: 1
  learning_rate: 0.0002
  num_train_epochs: 2
  lr_scheduler_type: cosine
  warmup_steps: 16
  optim: adamw_8bit
  metric_for_best_model: eval_loss
  load_best_model_at_end: true
  save_total_limit: 2
  seed: 42
  remove_unused_columns: false
  max_grad_norm: 1.0
  weight_decay: 0.08
  torch_compile: false
```