ruGPT-3.5-13B / Saiga2

LoRA адаптер для ruGPT3.5-13B обученный на коллекции датасетов Saiga.

Конфигурация: https://github.com/EvilFreelancer/impruver/blob/main/recipes/configs/ruGPT-3.5/13B_lora_saiga2.yaml

Адаптер обучался на 1x RTX 4090, для этого потребовалось примерно 18.2Gb VRAM и заняло 16h 58m.

output_dir: ./models/ruGPT35_13B_lora_saiga2
train_path: ./train.ruGPT35_13B_saiga2.jsonl
val_path: ./val.ruGPT35_13B_saiga2.jsonl

datasets:
  - name: IlyaGusev/ru_turbo_alpaca
    converter: impruver.instruction_to_messages
  - name: IlyaGusev/ru_turbo_alpaca_evol_instruct
    converter: impruver.instruction_to_messages
  - name: IlyaGusev/ru_turbo_saiga
    converter: impruver.dialog_to_messages
  - name: IlyaGusev/ru_sharegpt_cleaned
    converter: impruver.dialog_to_messages
  - name: IlyaGusev/oasst1_ru_main_branch
    converter: impruver.dialog_to_messages
  - name: lksy/ru_instruct_gpt4
    converter: impruver.converters.instruction_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: 1024

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: 128
  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