Datasets:
Add Dataset #3: personahub_code_v2_34999 (Tulu-3 SFT subset) (+2,584 records)
Browse files34,999 rows inspected row-by-row; 3,620 execution-verifiable via prompt-example
derivation (new Pattern I: input/output pair); 2,832 passed sandbox execution;
-245 benchmark-overlap, -3 name-mismatch, 0 flaky on determinism double-run;
2,584 accepted. License recorded as ODC-BY-1.0 (INFERRED, Tulu-3 family).
Total repo: 13,532 records, 100% execution-verified, 100% <=16k tokens.
Logs/reports/card/inventory updated; audit trail 123,944 log rows.
Dataset #4 (m-a-p/Code-Feedback) formally DEFERRED with documented rationale.
- 16k_context/train.jsonl +2 -2
- README.md +7 -3
- metadata/dataset_inventory.json +4 -4
- metadata/processing_log.jsonl +2 -2
- metadata/rejected_examples.jsonl +2 -2
- metadata/token_manifest.jsonl +0 -0
- reports/dataset_quality_report.md +35 -3
- reports/verification_report.md +14 -2
16k_context/train.jsonl
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8d9fc24fd4c947ecafd53349eaa9140437c9522812367766fc746bafaf2689b5
|
| 3 |
+
size 53465453
|
README.md
CHANGED
|
@@ -23,9 +23,9 @@ Priority order used for curation: **quality > correctness > usefulness > diversi
|
|
| 23 |
## Contents
|
| 24 |
|
| 25 |
```
|
| 26 |
-
├── 16k_context/train.jsonl #
|
| 27 |
├── over_16k/train.jsonl # empty — no accepted record exceeded 16k tokens
|
| 28 |
-
├── metadata/ # access check, dataset inventory, per-row processing log (
|
| 29 |
└── reports/ # dataset quality report + verification report (methods, evidence, limitations)
|
| 30 |
```
|
| 31 |
|
|
@@ -71,7 +71,11 @@ All records derive from execution-verified, training-intent, permissively licens
|
|
| 71 |
|
| 72 |
- Upstream pipeline **Magicoder / OSS-Instruct** (Wei et al., ICML 2024, [arXiv:2312.02120](https://arxiv.org/abs/2312.02120)); solutions generated by gpt-3.5-turbo-1106 from real OSS seeds with test generation + execution filtering documented in the paper; released file is the "decontaminated" variant. Released rows do not include test artifacts, so every Magicoder record here was accepted only after passing tests derived by us from the problem's own concrete examples (or the row's released/inline checks).
|
| 73 |
|
| 74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
Please attribute the upstream datasets and papers above when using this data. This repository's additions (curation, transformations, metadata) are provided under Apache-2.0.
|
| 77 |
|
|
|
|
| 23 |
## Contents
|
| 24 |
|
| 25 |
```
|
| 26 |
+
├── 16k_context/train.jsonl # 13,532 records, each ≤ 16,000 tokens (o200k_base)
|
| 27 |
├── over_16k/train.jsonl # empty — no accepted record exceeded 16k tokens
|
| 28 |
+
├── metadata/ # access check, dataset inventory, per-row processing log (123,944 rows), rejection audit trail (110,412), token manifest
|
| 29 |
└── reports/ # dataset quality report + verification report (methods, evidence, limitations)
|
| 30 |
```
|
| 31 |
|
|
|
|
| 71 |
|
| 72 |
- Upstream pipeline **Magicoder / OSS-Instruct** (Wei et al., ICML 2024, [arXiv:2312.02120](https://arxiv.org/abs/2312.02120)); solutions generated by gpt-3.5-turbo-1106 from real OSS seeds with test generation + execution filtering documented in the paper; released file is the "decontaminated" variant. Released rows do not include test artifacts, so every Magicoder record here was accepted only after passing tests derived by us from the problem's own concrete examples (or the row's released/inline checks).
|
| 73 |
|
| 74 |
+
**Source 3 — 2,584 records:** [ai2-adapt-dev/personahub_code_v2_34999](https://huggingface.co/datasets/ai2-adapt-dev/personahub_code_v2_34999) (Tulu-3 SFT-mix subset; license **ODC-BY-1.0 inferred** from the documented Tulu-3 data family license — the mirror itself carries no license tag):
|
| 75 |
+
|
| 76 |
+
- Persona-driven code prompts from the **Tulu 3** pipeline ([arXiv:2411.15124](https://arxiv.org/abs/2411.15124)); responses Llama-3.1-405B-generated with **no released per-row verification**. Every accepted record passed tests derived by us from the prompt's own stated Input/Output examples, executed twice (determinism). Rows whose examples contradicted their solutions (a documented weakness of this source) were rejected automatically by that mechanism.
|
| 77 |
+
|
| 78 |
+
Curation performed for this repository (full details in `reports/`): 123,944 source rows inspected row-by-row across the three datasets → static gates, AST analysis, conservative test derivation from stated examples, isolated-sandbox re-execution of every accepted record (twice; determinism double-run for Sources 2–3), benchmark-decontamination screening vs HumanEval/MBPP/BigCodeBench/DS-1000 (−1,726 rows), deduplication (incl. cross-dataset code-hash dedup), minor repairs (7 rows), instruction-leak removal, rationale cleanup, schema normalization, and programmatic token counting (tiktoken `o200k_base`). Final: **13,532 records, 100% execution-verified as shipped, 100% ≤16k tokens**.
|
| 79 |
|
| 80 |
Please attribute the upstream datasets and papers above when using this data. This repository's additions (curation, transformations, metadata) are provided under Apache-2.0.
|
| 81 |
|
metadata/dataset_inventory.json
CHANGED
|
@@ -66,7 +66,7 @@
|
|
| 66 |
"name": "ai2-adapt-dev/personahub_code_v2_34999",
|
| 67 |
"url": "https://huggingface.co/datasets/ai2-adapt-dev/personahub_code_v2_34999",
|
| 68 |
"platform": "Hugging Face (AI2 mirror of Tulu-3 data)",
|
| 69 |
-
"license": "
|
| 70 |
"rows": 34999,
|
| 71 |
"size_bytes": 48449834,
|
| 72 |
"python_relevance": "HIGH (mostly Python; 'Write a python function ...')",
|
|
@@ -90,7 +90,7 @@
|
|
| 90 |
"limitations": "Sampled first row is internally inconsistent (example claims 7 unique words while listing 8) -> per-row verification mandatory; license unclear on mirror",
|
| 91 |
"tier": "3 (Usable after substantial filtering)",
|
| 92 |
"tier_reason": "Training intent and scale, but no released tests, unknown license tag on mirror, quality issues found in sampling",
|
| 93 |
-
"decision": "
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"name": "m-a-p/Code-Feedback",
|
|
@@ -342,13 +342,13 @@
|
|
| 342 |
"rank": 3,
|
| 343 |
"dataset": "ai2-adapt-dev/personahub_code_v2_34999",
|
| 344 |
"tier": 3,
|
| 345 |
-
"note": "
|
| 346 |
},
|
| 347 |
{
|
| 348 |
"rank": 4,
|
| 349 |
"dataset": "m-a-p/Code-Feedback (python subset)",
|
| 350 |
"tier": 3,
|
| 351 |
-
"note": "
|
| 352 |
}
|
| 353 |
]
|
| 354 |
}
|
|
|
|
| 66 |
"name": "ai2-adapt-dev/personahub_code_v2_34999",
|
| 67 |
"url": "https://huggingface.co/datasets/ai2-adapt-dev/personahub_code_v2_34999",
|
| 68 |
"platform": "Hugging Face (AI2 mirror of Tulu-3 data)",
|
| 69 |
+
"license": "ODC-BY-1.0 INFERRED (Tulu-3 family license documented for allenai/tulu-3-* releases; this ai2-adapt-dev mirror carries no license tag)",
|
| 70 |
"rows": 34999,
|
| 71 |
"size_bytes": 48449834,
|
| 72 |
"python_relevance": "HIGH (mostly Python; 'Write a python function ...')",
|
|
|
|
| 90 |
"limitations": "Sampled first row is internally inconsistent (example claims 7 unique words while listing 8) -> per-row verification mandatory; license unclear on mirror",
|
| 91 |
"tier": "3 (Usable after substantial filtering)",
|
| 92 |
"tier_reason": "Training intent and scale, but no released tests, unknown license tag on mirror, quality issues found in sampling",
|
| 93 |
+
"decision": "PROCESSED (rank #3) - 34,999 rows inspected; 2,584 accepted (individually verified via prompt-derived tests, double-run)"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"name": "m-a-p/Code-Feedback",
|
|
|
|
| 342 |
"rank": 3,
|
| 343 |
"dataset": "ai2-adapt-dev/personahub_code_v2_34999",
|
| 344 |
"tier": 3,
|
| 345 |
+
"note": "PROCESSED 2026-10-03: 34,999 rows inspected; 3,620 execution-verifiable; 2,832 passed; -245 benchmark-overlap, -3 name-mismatch; 2,584 accepted"
|
| 346 |
},
|
| 347 |
{
|
| 348 |
"rank": 4,
|
| 349 |
"dataset": "m-a-p/Code-Feedback (python subset)",
|
| 350 |
"tier": 3,
|
| 351 |
+
"note": "DEFERRED (not processed): multilingual chat-feedback format with no per-row tests; python subset queries are conversation-dependent (weak derivability); 5.2GB; low expected yield at our acceptance bar - documented decision per quality-first mandate"
|
| 352 |
}
|
| 353 |
]
|
| 354 |
}
|
metadata/processing_log.jsonl
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:734526dbbf72bb809ecaef62199f987f6edc4fb16e6588626a1d7b2f9e9ab2e4
|
| 3 |
+
size 50212320
|
metadata/rejected_examples.jsonl
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7e0272ab34a13d1eab5fe3cefab9cadc34c8b0531456a5d09095dc28c1c0a539
|
| 3 |
+
size 58015045
|
metadata/token_manifest.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
reports/dataset_quality_report.md
CHANGED
|
@@ -103,9 +103,41 @@ Processed only after Dataset #1 was fully completed and uploaded (sequential-wor
|
|
| 103 |
- Repairs: 0 · Cross-dataset duplicates removed: included in the 120 (code-hash match vs Dataset #1 or intra-Magicoder)
|
| 104 |
- Manual review: 20-row random sample read end-to-end; findings consistent with automated gates (demo `print` blocks in a few solutions execute harmlessly; retained and observable in `code`)
|
| 105 |
|
| 106 |
-
###
|
| 107 |
-
|
| 108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 109 |
|
| 110 |
|
| 111 |
|
|
|
|
| 103 |
- Repairs: 0 · Cross-dataset duplicates removed: included in the 120 (code-hash match vs Dataset #1 or intra-Magicoder)
|
| 104 |
- Manual review: 20-row random sample read end-to-end; findings consistent with automated gates (demo `print` blocks in a few solutions execute harmlessly; retained and observable in `code`)
|
| 105 |
|
| 106 |
+
## 3c. Dataset #3 processing summary — ai2-adapt-dev/personahub_code_v2_34999 (Tulu-3 SFT subset)
|
| 107 |
+
|
| 108 |
+
Processed only after Dataset #2 was fully completed and uploaded (sequential-workflow rule).
|
| 109 |
+
|
| 110 |
+
### Why processed despite Tier 3 (evidence & license diligence)
|
| 111 |
+
- **DOCUMENTED**: this exact subset (34,999 rows) is the personahub-code component of the official Tulu-3 SFT mix (arXiv:2411.15124); prompts+responses generated with Llama-3.1-405B-Instruct; training intent explicit.
|
| 112 |
+
- **INFERRED license**: the ai2-adapt-dev mirror carries no license tag; the Tulu-3 data family is released under ODC-BY-1.0 (documented on allenai/tulu-3-* cards). Recorded as INFERRED in every record's provenance (`license_note`) — attribution preserved.
|
| 113 |
+
- **VERIFIED_BY_US (negative evidence)**: sampled rows contain internally inconsistent examples → per-row verification mandatory; our derive-and-execute acceptance mechanism automatically rejects such rows (674 assertion failures among executed rows).
|
| 114 |
+
|
| 115 |
+
### Funnel (all 34,999 rows inspected row-by-row)
|
| 116 |
+
| Stage | Rows |
|
| 117 |
+
|---|---|
|
| 118 |
+
| Source rows (train split) | 34,999 |
|
| 119 |
+
| Static gates: deps, prose-only/no-code answers, syntax, duplicates (**incl. cross-dataset code-hash dedup vs Datasets #1–#2**) | 32,290 pass (252 dep · 168 no-solution · 2,289 duplicate) |
|
| 120 |
+
| Execution-verifiable via prompt-derived tests (Patterns A/C/E/F/G + new **Pattern I**: "for the input `LIT` … expected output is `LIT2`" with arity-checked single-def solutions) | 3,620 |
|
| 121 |
+
| No high-confidence derivable tests | 28,670 → rejected_unverifiable |
|
| 122 |
+
| Sandbox execution | 2,832 / 3,620 passed (78.2%) |
|
| 123 |
+
| Benchmark screening (when-in-doubt-exclude) | −245 rejected_benchmark |
|
| 124 |
+
| Post-execution gates: name mismatch (3), determinism double-run (0 flaky) | −3 |
|
| 125 |
+
| **Final accepted (Dataset #3)** | **2,584** |
|
| 126 |
+
|
| 127 |
+
### Per-§21 counters (Dataset #3)
|
| 128 |
+
- Rows inspected: 34,999 (100%) · accepted: 2,584 · rejected: 32,415 — unverifiable 29,458 · duplicate 2,289 · dependency 252 · benchmark 245 · low_quality 171
|
| 129 |
+
- Execution-verified as shipped: 2,584/2,584, each passed twice (selection + determinism double-run)
|
| 130 |
+
- Tests: all derived from prompt-stated examples (majority 1 assert from the Input/Output pair; tagged in `example_quality`)
|
| 131 |
+
- Token counts (o200k_base): min 926 · p50 1,090 · max 1,882 → all ≤16K; over_16k additions: 0
|
| 132 |
+
- Manual review: 14-row random sample read end-to-end; derived tests faithful to prompt examples; persona-flavored instructions tagged (`persona_flavored_instruction`); reasoning mostly factual AST summaries (responses are code-only) — honest, non-fabricated
|
| 133 |
+
- Known weakness (documented): single-example verification is thinner than multi-assert suites; `example_quality` tags expose assert counts per record so downstream users can filter (e.g., ≥2 asserts) if desired
|
| 134 |
+
|
| 135 |
+
### Combined repository totals after Datasets #1+#2+#3
|
| 136 |
+
- **13,532 accepted records** (10,275 + 673 + 2,584), 100% execution-verified as shipped, 100% ≤16K tokens, 0 duplicate code hashes across sources.
|
| 137 |
+
- Audit trail: 123,944 processing-log rows; 110,412 rejected-row records with reasons.
|
| 138 |
+
|
| 139 |
+
### Dataset #4 decision (m-a-p/Code-Feedback): DEFERRED — not processed
|
| 140 |
+
Documented rationale: multilingual chat-feedback format without per-row tests; the python-subset queries are conversation-dependent (weak example derivability at our acceptance bar); 5.2GB payload; expected verified yield low. Per the quality-first mandate ("do not force every dataset into a tier"; small suitable set acceptable), processing stopped at three datasets.
|
| 141 |
|
| 142 |
|
| 143 |
|
reports/verification_report.md
CHANGED
|
@@ -37,6 +37,15 @@ Every candidate row's code was treated as **untrusted** and executed as follows:
|
|
| 37 |
| **Determinism double-run** | every candidate final artifact executed twice; inconsistent → dropped | 675 | 673 | 2 flaky dropped |
|
| 38 |
| Final invariant | shipped Magicoder records | **673** | **673 (100%)** | 0 |
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
### Failure taxonomy
|
| 41 |
- Core 13 failures: 6 solution-contradicts-own-test → `rejected_incorrect` (incl. a genuinely wrong `is_prime(9)`, a duplicate-removal bug, a `RuntimeError: dictionary keys changed during iteration`); 7 FileNotFound → repaired with fixtures built from instruction-described file contents → `accepted_after_fix`.
|
| 42 |
- Derived 597 failures: 437 AssertionError (solution inconsistent with the instruction's stated example, or conservative mis-extraction), 35 FileNotFound (file-based tasks without derivable fixtures), 34 NameError (instruction function name ≠ solution def), 27 TypeError, 63 other, 1 ImportError → all `rejected_unverifiable` (we do not claim they are incorrect when the derivation itself could be the failing part — honest labeling).
|
|
@@ -70,7 +79,8 @@ Results over the 11,725-row accepted pool (Dataset #1):
|
|
| 70 |
- HumanEval/MBPP screen: 842 auto-reject + 563 review-band → **1,405 rejected_benchmark** (typical hits: instruction ≡ MBPP task "two lists share a common element", "largest prime factor", "longest increasing subsequence length", "gcd of two positive integers", "difference between sum of squares and square of sum").
|
| 71 |
- BigCodeBench/DS-1000 screen: 2 auto-reject + 3 review-band → **5 rejected_benchmark**.
|
| 72 |
- Dataset #2 (Magicoder) 746-row passed pool: **71 rejected_benchmark** (e.g. "find all prime numbers up to a given n" ≡ MBPP task; review-band folded into rejections under the same when-in-doubt policy).
|
| 73 |
-
-
|
|
|
|
| 74 |
- Not screened: LiveCodeBench (recent contest problems; our sources are OSS-function-derived — different provenance class; documented as residual risk), SWE-bench family (repo-level issues; structurally absent from these sources).
|
| 75 |
|
| 76 |
## 5. Token accounting (§12)
|
|
@@ -79,15 +89,17 @@ Results over the 11,725-row accepted pool (Dataset #1):
|
|
| 79 |
- Counted: the **complete serialized JSON record** (instruction + reasoning + code + tests + documentation + verification + quality + provenance).
|
| 80 |
- Result (Dataset #1): min 728 · p50 1,009 · mean 1,038 · p95 1,328 · max 2,675.
|
| 81 |
- Result (Dataset #2): min 891 · p50 1,159 · max 2,391.
|
|
|
|
| 82 |
- Combined repository: **0 records > 16,000 tokens**; `over_16k/train.jsonl` is intentionally empty.
|
| 83 |
|
| 84 |
## 6. Manual review (§8 individual assessment)
|
| 85 |
|
| 86 |
-
A stratified random sample of **40 accepted records (Dataset #1)**
|
| 87 |
- No fabricated or inconsistent reasoning observed; extracted rationales match the verified code.
|
| 88 |
- 2 systemic issues found and fixed globally: (a) instruction↔solution **function-name mismatch** → 33 rows dropped (`rejected_low_quality`); (b) **transition-sentence residue** in reasoning/documentation ("Here is the implementation:") → cleaned in 4,943 records; 129 thin rationales replaced with factual AST-based summaries (tagged `reasoning_generated_factual_summary_from_code`; no fake monologue).
|
| 89 |
- Minor, accepted-and-documented observations: 1 record retains a module-level demo `print` (executes harmlessly); 2 records have loose spec naming ("extended Euclidean" implemented as plain Euclid; out-of-range behavior via IndexError instead of explicit raise) — both fully consistent with their stated tests.
|
| 90 |
- Dataset #2 sample: consistent with gates; several solutions retain upstream demo blocks (`# Test the function ... print(...)`) which execute harmlessly during verification and remain visible in `code` (no hidden behavior); Pattern-G prelude tests faithfully reproduce the problem's stated example values.
|
|
|
|
| 91 |
|
| 92 |
## 7. What we claim vs. what we don't (§22)
|
| 93 |
|
|
|
|
| 37 |
| **Determinism double-run** | every candidate final artifact executed twice; inconsistent → dropped | 675 | 673 | 2 flaky dropped |
|
| 38 |
| Final invariant | shipped Magicoder records | **673** | **673 (100%)** | 0 |
|
| 39 |
|
| 40 |
+
### Dataset #3 stages (ai2-adapt-dev/personahub_code_v2_34999)
|
| 41 |
+
|
| 42 |
+
| Stage | What ran | Rows | Passed | Failed |
|
| 43 |
+
|---|---|---|---|---|
|
| 44 |
+
| Parse & compile (AST) | assistant responses parsed; fenced/raw code AST-validated | 34,999 | 32,290 gate-pass | dep/no-solution/dup/syntax rejections |
|
| 45 |
+
| Sandbox execution | solution + tests derived from prompt-stated examples (Patterns A/C/E/F/G/I) | 3,620 | 2,832 (78.2%) | 788 (674 assertion — incl. the source's documented internally-inconsistent examples —, 53 name, 26 type, 5 file, 30 other) |
|
| 46 |
+
| **Determinism double-run** | every candidate final artifact executed twice | 2,587 | 2,584 | 0 flaky (+3 name-mismatch dropped pre-run) |
|
| 47 |
+
| Final invariant | shipped personahub records | **2,584** | **2,584 (100%)** | 0 |
|
| 48 |
+
|
| 49 |
### Failure taxonomy
|
| 50 |
- Core 13 failures: 6 solution-contradicts-own-test → `rejected_incorrect` (incl. a genuinely wrong `is_prime(9)`, a duplicate-removal bug, a `RuntimeError: dictionary keys changed during iteration`); 7 FileNotFound → repaired with fixtures built from instruction-described file contents → `accepted_after_fix`.
|
| 51 |
- Derived 597 failures: 437 AssertionError (solution inconsistent with the instruction's stated example, or conservative mis-extraction), 35 FileNotFound (file-based tasks without derivable fixtures), 34 NameError (instruction function name ≠ solution def), 27 TypeError, 63 other, 1 ImportError → all `rejected_unverifiable` (we do not claim they are incorrect when the derivation itself could be the failing part — honest labeling).
|
|
|
|
| 79 |
- HumanEval/MBPP screen: 842 auto-reject + 563 review-band → **1,405 rejected_benchmark** (typical hits: instruction ≡ MBPP task "two lists share a common element", "largest prime factor", "longest increasing subsequence length", "gcd of two positive integers", "difference between sum of squares and square of sum").
|
| 80 |
- BigCodeBench/DS-1000 screen: 2 auto-reject + 3 review-band → **5 rejected_benchmark**.
|
| 81 |
- Dataset #2 (Magicoder) 746-row passed pool: **71 rejected_benchmark** (e.g. "find all prime numbers up to a given n" ≡ MBPP task; review-band folded into rejections under the same when-in-doubt policy).
|
| 82 |
+
- Dataset #3 (personahub) 2,832-row passed pool: **245 rejected_benchmark** (predominantly MBPP generic-task containment).
|
| 83 |
+
- Total excluded for benchmark overlap across all datasets: **1,726**.
|
| 84 |
- Not screened: LiveCodeBench (recent contest problems; our sources are OSS-function-derived — different provenance class; documented as residual risk), SWE-bench family (repo-level issues; structurally absent from these sources).
|
| 85 |
|
| 86 |
## 5. Token accounting (§12)
|
|
|
|
| 89 |
- Counted: the **complete serialized JSON record** (instruction + reasoning + code + tests + documentation + verification + quality + provenance).
|
| 90 |
- Result (Dataset #1): min 728 · p50 1,009 · mean 1,038 · p95 1,328 · max 2,675.
|
| 91 |
- Result (Dataset #2): min 891 · p50 1,159 · max 2,391.
|
| 92 |
+
- Result (Dataset #3): min 926 · p50 1,090 · max 1,882.
|
| 93 |
- Combined repository: **0 records > 16,000 tokens**; `over_16k/train.jsonl` is intentionally empty.
|
| 94 |
|
| 95 |
## 6. Manual review (§8 individual assessment)
|
| 96 |
|
| 97 |
+
A stratified random sample of **40 accepted records (Dataset #1)**, **20 (Dataset #2)**, and **14 (Dataset #3)** was read end-to-end (instruction ↔ reasoning ↔ code ↔ tests ↔ documentation consistency). Findings:
|
| 98 |
- No fabricated or inconsistent reasoning observed; extracted rationales match the verified code.
|
| 99 |
- 2 systemic issues found and fixed globally: (a) instruction↔solution **function-name mismatch** → 33 rows dropped (`rejected_low_quality`); (b) **transition-sentence residue** in reasoning/documentation ("Here is the implementation:") → cleaned in 4,943 records; 129 thin rationales replaced with factual AST-based summaries (tagged `reasoning_generated_factual_summary_from_code`; no fake monologue).
|
| 100 |
- Minor, accepted-and-documented observations: 1 record retains a module-level demo `print` (executes harmlessly); 2 records have loose spec naming ("extended Euclidean" implemented as plain Euclid; out-of-range behavior via IndexError instead of explicit raise) — both fully consistent with their stated tests.
|
| 101 |
- Dataset #2 sample: consistent with gates; several solutions retain upstream demo blocks (`# Test the function ... print(...)`) which execute harmlessly during verification and remain visible in `code` (no hidden behavior); Pattern-G prelude tests faithfully reproduce the problem's stated example values.
|
| 102 |
+
- Dataset #3 sample: derived tests faithful to prompt Input/Output examples; persona-flavored instructions retained and tagged; reasoning fields are mostly factual AST summaries (source responses are code-only) — honest, never fabricated; single-assert verification thickness disclosed per-record via `example_quality` tags.
|
| 103 |
|
| 104 |
## 7. What we claim vs. what we don't (§22)
|
| 105 |
|