Skip to content

Function-level definitions for reports; local-LLM stage verdict - #138

Merged
ParsaVictor merged 1 commit into
mainfrom
phase-o/function-level
Oct 9, 2026
Merged

ParsaVictor merged 1 commit into
mainfrom
phase-o/function-level

Conversation

@ParsaVictor

Copy link
Copy Markdown
Owner
  • packet --json lists ranked definitions (path, Class.method, lines) for long reports — function-level localisation from the existing definition ranking; file ranking unchanged.
  • Function-level evaluation scripts (LocAgent metric). dev-fast func Acc@5/10: 0.229/0.292 (no file prior yet).
  • Local LLM stage (Qwen2.5-Coder-3B on CPU) rejected with numbers: copies our order, or judges worse without it (Acc@1 0.200 → 0.067). Script kept for API models.

Fourteen benchmark sets unchanged.

🤖 Generated with Claude Code

… stage verdict

- chunk_rank keeps each chunk's definition name and lines; `rank_definitions`
  returns them by score; `packet --json` lists them for a long report
  (`definitions`, 20 deep, NM_DEFINITIONS for research runs). File ranking
  unchanged (same scores; dev-fast lists identical).
- Function-level scoring scripts (gold from the patch at the base commit, as
  LocAgent defines it): dev-fast func Acc@5/10 0.229/0.292 before any file prior.
- Local LLM stage (Qwen2.5-Coder-3B, llama.cpp, CPU): copies the engine's order
  when shown it, judges worse when not (dev Acc@1 0.200 -> 0.067). Rejected for
  small local models; `llm_localize.py --shuffle` kept for API models.

Fourteen sets unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@ParsaVictor
ParsaVictor merged commit 6cb01a4 into main Oct 9, 2026
3 checks passed
@ParsaVictor
ParsaVictor deleted the phase-o/function-level branch October 9, 2026 11:16
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant