Tool use distillation experiments

[aryaprabhudesai]

Description

Adds a full distillation pipeline to improve Qwen3.5-9B's tool-use competency in CooperBench's multi-agent team setting. The pipeline downloads the CooperBench/team-coop dataset, converts teacher trajectories to the training format, runs SFT on Modal, uploads the adapter to HuggingFace, and serves the model via vLLM for evaluation.

Changes

• Dataset download & analysis (scripts/distill/download_team_coop.py, extract_successful.py, analyze_coverage.py): downloads CooperBench/team-coop from HuggingFace, filters to successful+verified trajectories, and reports scenario coverage across the 8 canonical team-harness tool patterns
• Format conversion (scripts/distill/convert_codex_to_miniswe.py): converts 1,153 Codex codex_stream_log teacher trajectories (gpt-5.5-hao) to mini_swe_agent_v2 format so training data matches inference format — handles [command] parsing, timing-line stripping, diff block removal, and exit-code extraction
• SFT training (scripts/distill/train_modal.py): QLoRA (4-bit NF4, r=64/α=128) on Qwen3.5-9B over 1,230 trajectories on a Modal A100-80GB; uses action masking, gradient checkpointing, and a volume-commit callback to persist checkpoints; downloads weights locally after training
• Adapter upload (scripts/distill/upload_adapter.py): uploads the LoRA adapter to CooperBench/qwen3.5-9b-tool-use-sft on HuggingFace
• vLLM serving (scripts/distill/serve_vllm_modal.py): serves base model + LoRA adapter on Modal via vllm==0.19.0 with qwen3_coder tool-call parser; eval agent config in agent_config_qwen35_sft.yaml

Checklist

• Code follows project style
• Tests pass locally
• Documentation updated (if needed)