seed: 0
wandb: false
output_dir: output/train
lora:
  r: 64
  alpha: 128
  dropout: 0.05
train:
  epochs: 1
  batch_size: 1
  grad_accum: 16
  lr: 0.0001
  lr_scheduler: cosine
  warmup_ratio: 0.05
  weight_decay: 0.01
  logging_steps: 5
  max_seq_len: 8192
  save_strategy: steps
  save_steps: 100
  save_total_limit: 2
  auto_resume: true
  token_budget: 8000
model: Qwen/Qwen3.6-27B
constitution: none; exact historical craft conversations reused without new constitution
  review
data_repo: dougalldeepmind/2026-09-15-nonmoral-original-7-mix
data_revision: b35ead8eaf4d59091e0ef1d08b187c7822f05632
base_model_revision: 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
allow_default_supervise: true
data_file: mixture.jsonl
profile:
  key: qwen36
  model: Qwen/Qwen3.6-27B
  match: qwen36
  family: Qwen3.6
  gpu:
    train: NVIDIA H200
    inference: NVIDIA H100 80GB HBM3
  thinking: true
  serving:
    max_num_seqs: 32
    reasoning_parser: qwen3
    tool_call_parser: qwen3_xml
    supports_prefix_caching: true
  template:
    assistant_header: '<|im_start|>assistant

      '
    turn_end: <|im_end|>
    prefill: '<think>

      '
    empty_think: '<think>


      </think>


      '
    think_close: </think>
    render_kwargs:
      preserve_thinking: true
  train:
    model_class: image_text_to_text
    load_in_4bit: false
    attn_implementation: sdpa
    lora_target_modules: model\.language_model\..*\.(q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj)$
    memory:
      H200:
        max_padded_tokens: 8000
        provenance: 'scratch/probe_batch_memory.py (commit 83343e7) on the 4xH200
          pod; Slack #fellows-only-callum 2026-08-08'