dsfasdfaserwer commited on
Commit
15388a6
·
verified ·
1 Parent(s): b945b03

revert: remove VALIDATION_LOG.md

Browse files
Files changed (1) hide show
  1. VALIDATION_LOG.md +0 -48
VALIDATION_LOG.md DELETED
@@ -1,48 +0,0 @@
1
- # VALIDATION LOG — batch 1 (2026-10-03)
2
-
3
- ## Environment
4
- Sandbox: Linux container, 1 GB RAM, Python 3 (tiktoken/datasets/pyarrow/sympy installed). No Docker.
5
- Executor: subprocess, RLIMIT_CPU 8s, wall timeout 10s, RLIMIT_AS 1.6GB, RLIMIT_NPROC 64, empty env, tmpdir cwd.
6
- Token counter: tiktoken o200k_base, limit 16,384 (instruction+cot+code+doc).
7
-
8
- ## Dataset processing order (top-tier first, per ranking)
9
- 1. **SWE-smith (#1)** — DEFERRED (not discarded): rows require per-repo Docker images
10
- (image_name field) to run fail→pass tests; no Docker in this sandbox. Infra phase later.
11
- Same deferral: R2E-Gym, SWE-Gym, OpenSWE, SWE-rebench, SWE-Fixer (repo environments needed).
12
- 2. **TACO (#2)** — processed: shard train-00002 (2,827 problems) streamed.
13
- Screening stats: {'examined': 398, 'no_tests': 421, 'no_python_sol': 197, 'starter_skipped': 82, 'all_failed': 109, 'passed': 289}
14
- -> 122 problems fully execution-verified into work/taco_pool.jsonl; 53 discarded (all 3 sampled
15
- solutions failed full-suite -> discard per max-2-attempt policy); 88 skipped (no Python solution),
16
- 31 skipped (starter_code/function-style, later batch), 173 skipped (<3 tests).
17
- 3. **NuminaMath-TIR (#32, Section 2 top)** — processed: full 72,441-row parquet streamed.
18
- Screening stats: {'examined': 219, 'no_code': 0, 'exec_fail': 21, 'answer_mismatch': 107, 'weak_reasoning': 1, 'passed': 90}
19
- -> 90 traces execution-verified (code ran + printed output matched boxed answer) into
20
- work/numina_pool.jsonl; 22 exec failures and 106 answer mismatches discarded.
21
- Environment fix during run: first pass ran with system python3 missing sympy/numpy (211 false
22
- exec-fails); re-run with sandbox python after pip install -> exec_fail dropped to 22.
23
- 4. Next queued: Code-Contests-Plus (#3), CodeContests (#4), APPS (#7), Eurus-2 code subset (#6),
24
- DeepCoder-Preview (#8), OpenCodeReasoning (#33), NuminaMath-TIR remainder.
25
-
26
- ## Selection (batch 1 = 19 examples)
27
- - TACO: 11 of 122 pool problems chosen for difficulty spread (EASY→MEDIUM_HARD), test-suite size
28
- (70–348 tests), and instructiveness. Trivial/near-duplicate problems skipped (worth-adding gate).
29
- All 11 source solutions were **rewritten idiomatically**; every rewrite re-verified against the
30
- FULL original suite: 166/166, 123/123, 348/348, 117/117, 131/131, 202/202, 132/132, 278/278,
31
- 146/146, 70/70, 81/81 = 1,794/1,794 test executions passed. 0 rewrites failed.
32
- - NuminaMath-TIR: 8 of 90 pool traces chosen for topic diversity (combinatorics, number theory,
33
- greedy/search, stars-and-bars, brute-force verification). All 8 re-executed after CRLF
34
- normalization; printed outputs matched boxed answers 8/8. Native CoT kept verbatim
35
- (cot_source=native_verified); too-simple traces (e.g. binary conversion) skipped by the gate.
36
- - Instructions authored per RULES.md §2 (realistic user voice, complete spec, no implementation
37
- leakage). CoTs authored per RULES.md §1 (edge cases, alternatives, self-correction, dry-runs,
38
- complexity). Docs per §3.
39
-
40
- ## Bucketing result
41
- under_16k: 19 (tokens 548–1,778) · over_16k: 0 · discarded at assembly: 0
42
- over_16k folder exists (data/over_16k/) and the bucketing gate is active; batch-1 sources simply
43
- did not produce >16k-token examples (expected later from repo-context/SWE-style sources).
44
-
45
- ## Rejection audit
46
- data/discarded/rejections_batch1.json + screening stats above (work/taco_stats.json,
47
- work/numina_stats.json). Screening rejections: TACO 465 rows skipped/failed, NuminaMath 129
48
- traces failed execution/answer checks.