Skip to content

1.4.0: traceback functions vote, function list for every prompt; phase-1 priors measured; fresh jsoup holdout confirms S1 - #144

Merged
ParsaVictor merged 5 commits into
mainfrom
s18/phase1-priors
Oct 9, 2026
Merged

ParsaVictor merged 5 commits into
mainfrom
s18/phase1-priors

Conversation

@ParsaVictor

Copy link
Copy Markdown
Owner

Session 18, phases 1 and 3. Every idea prototyped offline on SWE-bench dev first (both halves, bar fixed in advance: file Acc@1 or func Acc@5 ≥ +0.02 in dev-fast and dev-rest). Lite/Verified untouched.

Kept (engine)

  • Traceback → function vote: each frame (File "…", line N, in name / pytest path.py:N: in name) resolves to the indexed file (longest unique suffix) and the function it names (or the innermost one containing the line); test frames skipped; RRF weight 2 in definition_order (so MCP functions_to_look gets it too).
  • packet --json lists definitions for every prompt (short issues were automatic misses).
  • Definition ranker query-term cap kept the alphabetically first 256 terms → now title terms, then rarest, up to 1024.
dev (210 with function gold) func Acc@1 Acc@5 Acc@10
v1.3.0, fast / rest 0.135 / 0.166 0.269 / 0.299 0.346 / 0.369
this PR, fast / rest 0.173 / 0.196 0.308 / 0.329 0.365 / 0.399

File level unchanged on every dev instance (0.307/0.524/0.618/0.689).

Rejected with numbers (contributions-log §8.14–8.16)

commit-history prior (BugLocator: file@1 0.271→0.169), repro-code identifiers, test→module (fires on 0), RM3, doc→code bridge (ripgrep 0.500→0.375, axios 0.833→0.583), owner/unstemmed definition variants (NM_DIAG), dense short-list (near miss: file@1 +0.017/+0.024, +10.6 s/issue).

Holdout

Fresh concept-holdout5 (jsoup 1.18.1, Java, gold committed before any run): list R@3 v1.2.0 0.417 → v1.3.0/this 0.750, BM25 0.667 — S1 confirmed. Packet 0.667 / 0.235.

15 gold sets with this code: all equal to the numbers on record (dev 0.938, large 0.675, holdout 0.700, holdout-c 0.589, holdout-lang 0.589, holdout-ml 0.478, holdout-ml2 0.632, holdout-cfg 0.767, holdout-web 0.917/0.643, concept 0.679/0.392, ripgrep 0.500/0.156, click 0.958/0.342, cobra 0.833/0.357, axios 0.667/0.528) + jsoup 0.667/0.235.

Also: research tooling (priors.py, chunk_prf.py, chunk_ablate.py, dense_shortlist.py, doc_bridge.py, diag_eval.py, paper_tables.py), harness --no-bm25, llm_localize.py --model/--functions (ready; all five Baseten keys return HTTP 402 — no credit).

🤖 Generated with Claude Code

ParsaVictor and others added 5 commits October 9, 2026 20:48
…old locked before any run

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pt, NM_DIAG variants, prior prototypes

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ut confirms S1

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@ParsaVictor
ParsaVictor merged commit ee6cc27 into main Oct 9, 2026
3 checks passed
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