mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-07-28 13:26:48 +00:00
Andrew Park
5aa16473ff
Replace the ADP-relabel cold-start with a ToolScale-backed, execution-grounded rejection-sampling pipeline in the paper's real action space (arXiv:2511.21689). - expert_registry: faithful unified tool calling — one named tool per model (gpt_5, qwen3_32b, ...) + basic tools, per-instance subset/price sampling (§3.1/3.3) - toolorchestra/: split the 1663-line module into a package (prompts/experts/ sandbox/clients/parsing/workers/agent), behavior preserved - toolorchestra/rollout.py: faithful reason->act->observe loop (pure, injectable) - toolorchestra/unified.py: real teacher caller + tool dispatch - sft_data/toolscale.py: nvidia/ToolScale loader - sft_data/unified_serialize.py: trajectory -> conversations JSONL (<tool_call> tags) - sft_data/reject_sample.py: generate_sft_dataset (sample N -> verify -> keep) - scripts/orchestrator/build_unified_sft.py: CLI - 25 new offline tests; 127 orchestrator-learning tests green Verifier is a gold-action-coverage proxy (full ToolScale DB checker + LLM judge are follow-ups); deployed _run_rl still uses the 3-meta-tool eval port.