OpenCode
68a9071cf1
Local branch had diverged from origin/master (sibling commits on the same base). Rewrote local history linearly on top of origin/master, folding in all local content: exa/hot-platform discovery hunters, leaks ledger, nightly sweep orchestrator, updated .gitignore and skill docs, plus the latest hunt output and vault state. Remote-only files (vault keys, channel scripts) were restored rather than dropped, so the resulting tree is a full union of both sides.
1 line
2.2 KiB
JSON
1 line
2.2 KiB
JSON
{"repo": "aweai-team/aweagent", "files": {"configs/rollout/scale_swe.yaml": {"4d464fe96daac630168dd572b7cb95532403a6f7": {"t": 1786068416.39056, "n": 1950, "h": "ccda876e4c95dc22", "c": "# Example rollout config: collect teacher trajectories on ScaleSWE.\n#\n# A rollout config *is* the experiment — teacher endpoint, the system prompt /\n# skills under test, the dataset, samples per instance, and whether to verify.\n# To sweep prompts/teachers, copy this file and edit the marked lines; to reuse\n# one setup on different data, pass rollout(..., data_file=...).\n#\n# Usage:\n# from aweagent.server import rollout\n# res = await rollout(\n# \"configs/rollout/scale_swe.yaml\",\n# data_file=\"datasets/scale_swe/train.jsonl\", # overrides ${DATA_FILE}\n# )\n# print(res.run_dir, res.n_accepted, \"/\", res.n_trajectories, res.rollout_paths)\n\n# ── Teacher endpoint (any OpenAI-compatible API: DeepSeek, vLLM, sglang…) ──\nllm:\n backend: openai\n base_url: ${TEACHER_BASE_URL:-https://api.deepseek.com/v1}\n model: ${TEACHER_MODEL:-deepseek-chat}\n api_key: ${TEACHER_API_KEY}\n params:\n temperature: 1.0 # sampling (not eval's greedy 0); keep > 0 for num_rollouts > 1\n max_tokens: 4096\n\n# ── Harness under test — edit these to run a prompt experiment ──\nagent:\n type: search_swe\n max_steps: 200\n enable_search: false\n tool_call_format: codeact_xml\n # system_prompt_file: prompts/exp1_system.txt # ← swap the system prompt\n # skill_files: # ← attach solving skills\n # - skills/verifier_first.md\n\n# ── Dataset (data_file usually overridden per-run via rollout(data_file=...)) ──\ntask:\n type: scale_swe\n dataset_id: scale_swe\n data_file: ${DATA_FILE}\n\n# ── Verifier on → each trajectory gets an accepted/score verdict for\n# downstream reject-sampling. Set enabled: false to skip (no verdict). ──\neval:\n enabled: true\n timeout: 3600\n\nexecution:\n num_rollouts: 4 # samples per instance (needs temperature > 0 above)\n max_concurrent: 50\n output_path: ./results/rollout_server/scale_swe\n # save_trajectories is forced True by the rollout server regardless.\n"}}}} |