yurteg-qwen2.5-1.5b-sft

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.15.0

# ============================================================
# ЮрТэг — SFT Fine-Tuning Config для RunPod/Axolotl
# Модель: Qwen 2.5 1.5B Instruct
# Датасет: SuperPuperD/yurteg-legal-sft (635 train + 95 val)
# ============================================================

# --- Модель ---
base_model: Qwen/Qwen2.5-1.5B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# --- Квантизация (QLoRA 4-bit) ---
load_in_8bit: false
load_in_4bit: true

# --- Датасет ---
datasets:
  - path: SuperPuperD/yurteg-legal-sft
    type: chat_template
    chat_template: tokenizer_default
    field_messages: messages
    roles_to_train:
      - assistant
    train_on_eos: last
    data_files:
      - train.jsonl

# Валидация — 13% от train (≈82 записи, аналогично нашему val.jsonl)
val_set_size: 0.13

# --- Результат ---
output_dir: ./outputs/yurteg-qwen2.5-1.5b-sft

# --- LoRA (максимальное качество) ---
adapter: qlora
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true

# --- Длина и упаковка ---
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false

# --- Обучение (агрессивные настройки) ---
micro_batch_size: 2
gradient_accumulation_steps: 8
num_epochs: 8
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0001
warmup_ratio: 0.1
weight_decay: 0.01

# --- Точность ---
bf16: auto
tf32: true

# --- Оптимизация памяти ---
gradient_checkpointing: true
flash_attention: true

# --- Логирование и сохранение ---
logging_steps: 5
evals_per_epoch: 4
saves_per_epoch: 1
save_total_limit: 3

# --- Воспроизводимость ---
seed: 42

# --- Специальные токены (Qwen) ---
special_tokens:
  pad_token: "<|endoftext|>"

</details><br>

outputs/yurteg-qwen2.5-1.5b-sft

This model is a fine-tuned version of Qwen/Qwen2.5-1.5B-Instruct on the SuperPuperD/yurteg-legal-sft dataset. It achieves the following results on the evaluation set:

• Loss: 0.3193
• Ppl: 1.3762
• Memory/max Active (gib): 16.76
• Memory/max Allocated (gib): 16.76
• Memory/device Reserved (gib): 19.41

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 12
• training_steps: 128

Training results

Framework versions

• PEFT 0.18.1
• Transformers 5.3.0
• Pytorch 2.8.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2