Skip to content

Function level: classes out + file prior (dev func Acc@5 0.268 → 0.323); functions_to_look; Verified & ablation - #140

Merged
ParsaVictor merged 2 commits into
mainfrom
phase-o/function-level-2
Oct 9, 2026
Merged

ParsaVictor merged 2 commits into
mainfrom
phase-o/function-level-2

Conversation

@ParsaVictor

Copy link
Copy Markdown
Owner
  • Classes leave the function list (still vote for files); definitions fused with file rank. SWE-bench dev (189): func Acc@1/5/10 0.137/0.268/0.337 → 0.175/0.323/0.402.
  • MCP functions_to_look for pasted reports.
  • SWE-bench Verified (v1.2.0, once, 496/500): 0.405/0.669/0.732/0.804 vs BM25 0.216/0.391/0.490/0.641.
  • Ablation (Lite holdout 276): −definition ranking 0.388/0.569/0.652; −hygiene 0.467/0.707/0.743; packet only 0.366/0.536/0.558; full 0.486/0.710/0.750.

Fourteen benchmark sets unchanged.

🤖 Generated with Claude Code

ParsaVictor and others added 2 commits October 9, 2026 15:59
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
….323); functions_to_look over MCP; Verified + ablation measured

- A definition that contains others (a class) still votes for its file but
  leaves the function list; `definition_order` fuses the function ranking with
  each definition's file rank (Agentless's function step, without the LLM).
  SWE-bench dev, 189 issues: func Acc@1/5/10 0.137/0.268/0.337 -> 0.175/0.323/0.402.
- MCP: a pasted report also gets `functions_to_look` (5, `path:Class.method Lx-Ly`).
- Measured (v1.2.0 engine, run once): SWE-bench Verified 496/500 Acc@1/3/5/10
  0.405/0.669/0.732/0.804 vs BM25 0.216/0.391/0.490/0.641; ablation on the Lite
  holdout: without the definition ranking 0.388/0.569/0.652, without report
  hygiene 0.467/0.707/0.743, packet order only 0.366/0.536/0.558.

Fourteen sets unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@ParsaVictor
ParsaVictor merged commit 7106fd7 into main Oct 9, 2026
3 checks passed
@ParsaVictor
ParsaVictor deleted the phase-o/function-level-2 branch October 9, 2026 13:11
@ParsaVictor ParsaVictor mentioned this pull request Oct 9, 2026
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