_experiment_metadata:
  experiment_name: aimo3-exp-1
  timestamp: '2026-01-13T06:37:34.326062+08:00'
  wandb_run_id: 0pkyx3cq
experiment_name: aimo3-exp-1
exp_tag: sft-lora
framework: accelerate
method: sft
project:
  hf_username: vinoku89
  docker_username: vinoku89
  docker_image: vinoku89/llm-vastai-cuda-13.0.2-auto:latest
  git_repo: svk-123/distributed-sft-rl-inference-pipeline
  git_branch: dev
model:
  name: Qwen/Qwen3-0.6B
  use_quantization: false
  quantized_model: null
  trust_remote_code: true
data:
  handler: math_sft
  dataset_name: HAD653/GSM8K-OpenMath-MathReason-13k
  dataset_config: null
  train_file: null
  test_file: null
  max_seq_length: 8196
  problem_col: question
  solution_col: cot
  answer_col: final_answer
  use_thinking_format: true
  preprocessing:
    use_chat_template: true
training:
  output_dir: ./checkpoints
  num_train_epochs: -1
  max_steps: 10
  per_device_train_batch_size: 2
  per_device_eval_batch_size: 2
  gradient_accumulation_steps: 4
  learning_rate: 2.0e-05
  weight_decay: 0.01
  max_grad_norm: 1.0
  optim: paged_adamw_8bit
  lr_scheduler_type: linear
  warmup_ratio: 0.05
  warmup_steps: 5
  bf16: true
  fp16: false
  gradient_checkpointing: true
  gradient_checkpointing_kwargs:
    use_reentrant: true
  dataloader_pin_memory: true
  logging_steps: 10
  eval_strategy: epoch
  eval_steps: 500
  save_strategy: epoch
  save_steps: 500
  save_total_limit: 2
  seed: 42
  resume_from_checkpoint: null
  report_to: wandb
  model: Qwen/Qwen3-0.6B
  dataset: HAD653/GSM8K-OpenMath-MathReason-13k
  problem_col: question
  solution_col: cot
  max_seq_length: 8196
  epochs: -1
lora:
  enabled: true
  r: 16
  lora_alpha: 16
  lora_dropout: 0.05
  bias: none
  task_type: CAUSAL_LM
  target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj
prompts:
  system: You are a helpful and reflective maths assistant. Please reason step by
    step to put the answer in \boxed{}.
  user_suffix: '


    Please reason step by step, and put your final answer within \boxed{}.'
validation:
  enabled: true
  required_after_training: true
  datasets:
  - name: aimo3-ref-val-1
    path: vinoku89/aimo3-ref-val
    split: validation
    num_samples: 1
    metric: accuracy
    columns:
      question: problem
      ground_truth: answer
  - name: aimo3-syn-val-1
    path: vinoku89/aimo3-syn-val-1
    split: validation
    num_samples: 1
    metric: accuracy
    columns:
      question: problem
      ground_truth: answer
  generation:
    max_new_tokens: '[REDACTED]'
    temperature: 0.6
    top_p: 0.95
    top_k: 20
    do_sample: true
  thresholds: {}
  inference:
    engine: vllm
    tensor_parallel_size: 0
    batch_size: 8
    gpu_memory_utilization: 0.9
    max_model_len: 4096
    dtype: half
    max_num_seqs: 128
  wandb:
    enabled: true
    entity: null
  upload_results: true
  output_dir: validation_results
wandb:
  enabled: true
  tags:
  - model:qwen3-0.6b
  - dataset:gsm8k-openmath-mathreason-13k
  - training
huggingface:
  push_to_hub: true
  hub_model_id: vinoku89/aimo3-exp-1
  private: true
infrastructure:
  gpu: RTX_4090
  ngpu: 1
  disk: 100
  budget: 1.0
  timeout: 300
  auto_destroy: true
  watchdog:
    enabled: true
    grace_period: 600
    idle_threshold: 10
    idle_duration: 1200
    check_interval: 60
_runtime:
  accelerate_config: configs/accelerate_runtime.yaml
  deepspeed_config: configs/deepspeed_runtime.json
  deepspeed_preset: zero0
  num_gpus: 1
_training_timestamp: '20260113_063547'
_run_number: '003'