Files
hack/tools/scripts/llm-key-hunter/results/_cache/content/bagadir_eidon-swebench.json
T

1 line
4.2 KiB
JSON
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
{"repo": "bagadir/eidon-swebench", "files": {"README.md": {"8c3133ff1c7d5cacb2f2be7e4eed9dacd4363751": {"t": 1785746596.4309294, "n": 1935, "h": "a0cee7423c7ef41a", "c": "# Eidon × DeepSeek-R1 on SWE-bench Verified\n\n**Eidon** is a local-first repository intelligence engine. This repo runs it against the [SWE-bench Verified](https://www.swebench.com/) benchmark (500 real GitHub issues) using DeepSeek's thinking-mode model.\n\n## How it works\n\n**4-stage pipeline per task:**\n\n```\nStage 1 — Encode eidon analyze --fresh → .eidon/encoding (L0–L3 graph)\nStage 2 — Localize deepseek-chat → top-5 files most likely to fix the bug\nStage 3 — Patch deepseek-reasoner → think step-by-step → unified diff\nStage 4 — Repair deepseek-reasoner → fix hunk-offset errors, re-apply\n```\n\nEidon's `encoding` gives the model a **compressed L0–L3 architecture map** of the repo (purpose → interface → usage → full source) before it writes a single line of diff. This is the key differentiator vs. agents that search files at random.\n\n## Key choices\n\n| Decision | Value | Why |\n|----------|-------|-----|\n| Localize | `deepseek-chat` | Fast, structured JSON, no reasoning overhead |\n| Patch | `deepseek-reasoner` | Thinking mode — explicit chain-of-thought on root cause |\n| Repair | `deepseek-reasoner` | Reasons about why the hunk failed |\n| Token budget | 32K (Eidon encoding) | Full architecture context |\n| Patch tokens | 8K | Reasoner needs room for thinking tokens |\n\n## Running locally\n\n```bash\nnpm install -g eidoncore\neidon activate \"$EIDON_LICENSE_KEY\"\npip install -r swebench/requirements.txt\nexport DEEPSEEK_API_KEY=sk-...\npython swebench/eidon_agent.py --tasks 10\n```\n\n## Running on GitHub Actions (500 tasks, 5 shards)\n\n1. Fork this repo\n2. Add secrets: `DEEPSEEK_API_KEY`, `EIDON_LICENSE_KEY`\n3. Push — the workflow triggers automatically\n\n**Estimated cost:** ~$7–10 for all 500 tasks at $0.28/1M input, $0.42/1M output.\n\n## Submitting\n\n```bash\npip install sb-cli\nsb-cli submit --predictions predictions.json --split verified\n```\n\n## Leaderboard tag\n\n`eidon-deepseek-reasoner`\n"}, "2e82d1b36268842ee9220194e164c3ac7aceae4c": {"t": 1785766175.5827298, "n": 1935, "h": "a0cee7423c7ef41a", "c": "# Eidon × DeepSeek-R1 on SWE-bench Verified\n\n**Eidon** is a local-first repository intelligence engine. This repo runs it against the [SWE-bench Verified](https://www.swebench.com/) benchmark (500 real GitHub issues) using DeepSeek's thinking-mode model.\n\n## How it works\n\n**4-stage pipeline per task:**\n\n```\nStage 1 — Encode eidon analyze --fresh → .eidon/encoding (L0–L3 graph)\nStage 2 — Localize deepseek-chat → top-5 files most likely to fix the bug\nStage 3 — Patch deepseek-reasoner → think step-by-step → unified diff\nStage 4 — Repair deepseek-reasoner → fix hunk-offset errors, re-apply\n```\n\nEidon's `encoding` gives the model a **compressed L0–L3 architecture map** of the repo (purpose → interface → usage → full source) before it writes a single line of diff. This is the key differentiator vs. agents that search files at random.\n\n## Key choices\n\n| Decision | Value | Why |\n|----------|-------|-----|\n| Localize | `deepseek-chat` | Fast, structured JSON, no reasoning overhead |\n| Patch | `deepseek-reasoner` | Thinking mode — explicit chain-of-thought on root cause |\n| Repair | `deepseek-reasoner` | Reasons about why the hunk failed |\n| Token budget | 32K (Eidon encoding) | Full architecture context |\n| Patch tokens | 8K | Reasoner needs room for thinking tokens |\n\n## Running locally\n\n```bash\nnpm install -g eidoncore\neidon activate \"$EIDON_LICENSE_KEY\"\npip install -r swebench/requirements.txt\nexport DEEPSEEK_API_KEY=sk-...\npython swebench/eidon_agent.py --tasks 10\n```\n\n## Running on GitHub Actions (500 tasks, 5 shards)\n\n1. Fork this repo\n2. Add secrets: `DEEPSEEK_API_KEY`, `EIDON_LICENSE_KEY`\n3. Push — the workflow triggers automatically\n\n**Estimated cost:** ~$7–10 for all 500 tasks at $0.28/1M input, $0.42/1M output.\n\n## Submitting\n\n```bash\npip install sb-cli\nsb-cli submit --predictions predictions.json --split verified\n```\n\n## Leaderboard tag\n\n`eidon-deepseek-reasoner`\n"}}}}