dedilute-patch-gate / gate /RESULTS.md
adtygan's picture
Upload gate/RESULTS.md with huggingface_hub
f48b3d4 verified
|
Raw
History Blame Contribute Delete
2.96 kB

§4 Week-1 Gate — Results Log

Auditable record of the GO/NO-GO gate. Every number traces to public data + a script in this folder.

Env note: datasets.load_dataset hangs in this box (it resolves blobs via the dead cdn-lfs.huggingface.co, HTTP 000). Workaround used everywhere: hf_hub_download of the parquet directly (Xet/CAS path, works), then read with pyarrow. HF token in gitignored ../../.env.


Step 2 — is the 32k+ slice real (in TOKENS, not chars)? ✅ PASS

step2_tokenize_packs.py --pack 40K --scope verified · tokenizer Qwen/Qwen3-Coder-30B-A3B-Instruct · data princeton-nlp/SWE-bench_bm25_40KSWE-bench_Verified (500) · CPU-only.

  • Token-length: p10 44,514 · p50 46,598 · p90 47,712 · p99 48,178.
  • 100% of the 500 clear 32k (and 24k, 16k).
  • char/token median 3.63 → raw HF char_len overstates length ~3.6×; tokenizing locally was necessary.
  • Length-vs-difficulty confound (2602.16069) absent: median tokens flat across difficulty (<15min=46,774 · 15min-1h=46,455 · 1-4h=46,517 · >4h=47,119).

Caveat: packs are uniformly ~46k because BM25 packs to a ~40k budget; "long" is a property of the retrieval construction, not of natural trajectory length. Sufficient for the dilution premise (context genuinely >32k), consistent with proposal §8 ("construct the long context yourself").

Artifacts: out/step2_40K_verified.json (per-instance), out/step2_summary_verified.json.

Status: 32k premise holds. Confound guarded. → proceed.


Step 1 — build oracle-collapsed (de-diluted span) ✅ DONE (CPU, free)

step1_build_oracle_collapsed.py --ctx 15 · gold files from princeton-nlp/SWE-bench_oracle, hunk ranges from gold patch (- side) · Verified 500 · tokenizer Qwen3-Coder.

C* = [ <issue> ; gold files collapsed to ±15 lines around each hunk ; <failing_test> ].

  • Token dist: p10 881 · p50 1,489 · p75 2,250 · p90 3,239 · p99 8,424.
  • ~31× smaller than the diluted pack (step-2 median 46,598). Even p99 (8.4k) is well below the ~46k dilution regime → the de-diluted span sits beneath dilution onset, as the premise requires.
  • 499/500 parsed clean; 1 edge case (astropy-13398: gold patch creates a file → no pre-image, 3/4 files still collapsed). Spot-check astropy-12907: ±15 window correctly contains the buggy line the gold patch fixes.

Artifact: out/step1_oracle_collapsed_ctx15.json (per-instance context + token count). Status: de-diluted span built and verified. → ready for steps 3/4 (need teacher on GPU 1 + scoring).

Step 1b (open infra for 3/4): teacher = Qwen3-Coder-30B-A3B on GPU 1 (vllm not yet installed);

scoring = proxy (gold-hunk localization) first, then hosted sb-cli for binding FAIL_TO_PASS.

Step 3 — per-span-type compressibility gap — TODO (existential kill criterion)

Step 4 — 3-condition scissor ceiling (trunc / mask / oracle) — TODO (existential kill criterion)