potteryrage/aileron-0.8-laminar-student-v1

Aileron 0.8 Laminar Student v1

This is a custom Laminar-native sparse student checkpoint distilled from Qwen/Qwen3.5-0.8B.

It is not a standard Hugging Face AutoModelForCausalLM checkpoint. The artifact contains the sparse student modules used to patch selected Qwen decoder MLPs at runtime:

• learned residency-constrained router
• page-local SwiGLU experts
• cache-resident correction path
• trainable final-head bias/scale

Files

• laminar_student.pt - trained sparse student checkpoint.
• training_config.json - full Modal training config.
• metadata.json - run metadata.
• laminar_student_metrics.csv - training metrics.
• eval_summary.json - final evaluation summary.
• eval_rollouts.csv - final rollout outputs and gate metrics.
• aileron_train_laminar_student.py - trainer used for this artifact.
• aileron_eval_laminar_student.py - evaluator used for the final gate.

Training Run

• GPU: Modal B200
• Steps: 20,000
• Sequence length: 128
• Sparse layers: 0,1,2
• AP/pages/expert hidden: AP32, 256 pages, H32
• Correction rank: 16
• Wall time: about 1h53m

Required Decoding Policy

The checkpoint passed the basic 256-token chat gate with:

{
  "repetition_penalty": 1.08,
  "no_repeat_ngram_size": 4,
  "temperature": 0.0
}

Do not evaluate this artifact with unconstrained greedy decoding; the raw greedy run still showed long-horizon repetition. The decoding policy above is part of the functional artifact contract.

Final Gate Summary

eval_summary.json:

• passed_basic_chat_gate: true
• max_repeated_4gram_rate_256: 0.0
• max_line_repeat_256: 1
• story_contamination: false
• malformed_fragment: false

This is a functional Laminar sparse-student checkpoint for continued runtime and MLX/Laminar integration work, not a general-purpose replacement for the dense Qwen teacher.