Physics-Synthetic-Data/Qwen2.5-coder-7B-caption2TikZ-SFT

# ---------------- 核心模型路径 ----------------
base_model: /nfs/data/johnsonk/mllm_models/RAFT/Oct2/lr1e-5_epoch4_warm0.03_training_6k_GPU_0_1_2_3_20251001_172635/trained_model   
# output_dir: /work/nvme/bckr/wli18/RAFT/exp_1
dataset_prepared_path: last_run_prepared
# --------------- 数据集 ----------------------
datasets:
  # 训练集
  - path: /home/johnsonk/RAFT/1exp/exp_1.0.2_questionize_high_school_physics_861_img/result/image_captions_2_sft_data_format_shuffled.jsonl
    type: chat_template
    split: train
    field_messages: messages
    message_property_mappings:
      role: role
      content: content
val_set_size: 0.05    

# --------------- 训练超参 ---------------------

sequence_len: 2048
micro_batch_size: 1 
gradient_accumulation_steps: 8
num_epochs: 3 
learning_rate: 1e-5
optimizer: adamw_torch
lr_scheduler: cosine
warmup_ratio: 0.03
weight_decay: 0.01



# max_steps: 100  # <--- 新增此行，强制运行100步
# --------------- 精度 / 显存 ------------------
bf16: auto
tf32: true
gradient_checkpointing: true
# gradient_checkpointing_kwargs:
#   use_reentrant: false

# --------------- DeepSpeed --------------------
deepspeed: /home/johnsonk/RAFT/configs/deepspeed_stage3.json

# ----------- 关闭所有 LoRA / 量化 -------------
# (配置已为全参数训练)
# load_in_4bit: true
# adapter: qlora
# lora_r: 16
# lora_alpha: 32
# lora_target_modules:
#   - q_proj
#   - k_proj
#   - v_proj
#   - o_proj
#   - down_proj
#   - up_proj
# lora_mlp_kernel: true
# lora_qkv_kernel: true
# lora_o_kernel: true

# -------------- 其它杂项 ----------------------
strict: false
chat_template: qwen_25
save_steps: 100000  # 每50个step保存一次

# [关键] 设置最多只保留2个最新的checkpoint，自动删除旧的
save_total_limit: 0

evals_per_epoch: 1
# saves_per_epoch: 100
logging_steps: 1
flash_attention: true


### WandB ###
report_to: wandb
wandb_project: test_axolotl
wandb_entity: johnson0213-ucla
wandb_name: Qwen2.5-Coder-7B-TikZ-SFT-860-v1-3epoch-grad_acc8-lr1e-5