model:
  base_model: deepseek-ai/deepseek-math-7b-instruct
  trust_remote_code: true
  use_bf16: true
  load_in_4bit: true
  bnb_4bit_quant_type: nf4
  bnb_4bit_use_double_quant: true
  attn_implementation: null
  lora:
    r: 64
    alpha: 128
    dropout: 0.05
    bias: none
    target_modules:
      - q_proj
      - k_proj
      - v_proj
      - o_proj
      - gate_proj
      - up_proj
      - down_proj

data:
  train_file: data/releases/v1/train.parquet
  validation_file: data/releases/v1/validation.parquet
  prompt_field: prompt
  target_field: target
  final_answer_field: final_answer
  proof_field: proof_formal
  max_seq_length: 2048
  max_train_samples: null
  max_eval_samples: 2000
  system_prompt: |
    You are a rigorous mathematical reasoning assistant specialized in unsolved
    conjectures. Produce clear, checkable reasoning and avoid claiming a full
    proof unless it is explicitly available in the task context.

training:
  output_dir: model_development/runs/deepseek-math-lora
  num_train_epochs: 1
  per_device_train_batch_size: 1
  per_device_eval_batch_size: 1
  gradient_accumulation_steps: 16
  learning_rate: 2.0e-5
  weight_decay: 0.01
  warmup_ratio: 0.03
  lr_scheduler_type: cosine
  max_grad_norm: 1.0
  gradient_checkpointing: true
  logging_steps: 10
  save_steps: 250
  eval_steps: 250
  save_total_limit: 3
  dataloader_num_workers: 2
  seed: 17

hub:
  push_to_hub: true
  repo_id: NorthernTribe-Research/math-conjecture-model
  private: false
  commit_message: Train DeepSeek-Math LoRA on conjecture corpus.

credentials:
  path: huggingface-api-key.json