aaryand/context-adherence-guardrail-10k

Context-Adherence Guardrail — training data (10,710)

Training data for a single-token RAG-groundedness guardrail. Each item is a (question, context, response) triple with a human- or construction-derived PASS/FAIL label under one Behavior Spec:

FAIL iff the response makes at least one factual claim unsupported by or contradicting the retrieved context — truth in the real world is irrelevant (strict grounding). PASS otherwise, including responses that decline to answer for lack of context.

Trained model: aaryand/qwen3-1.7b-context-adherence-guardrail (spec adherence 0.556 → 0.718 over the prompted base, on 8,295 human-annotated held-out items).

Fields

field	meaning
id	stable identifier (train-real-* / train-syn-*)
task_type	QA, Summary, or Data2txt
question	question, or null for summarisation/data-to-text
context	the retrieved context the response must be grounded in
response	the candidate answer being judged
label	PASS or FAIL
meta	provenance: component (real/synthetic), generator, source_id, and for synthetic items the grid cell and the injected claim

Splits

Nested, so a data-efficiency curve is a strict subset sweep: 250 ⊂ 1,000 ⊂ 4,000 ⊂ 10,710. Exactly 50/50 PASS/FAIL within each task type.

split	n	FAIL	synthetic
train_250	250	125	31
train_1000	1,000	501	97
train_4000	4,000	2,000	370
train (full)	10,710	5,355	1,009

Two things worth knowing before you use this

1. ~91% is human-labelled, not distilled. The bulk is the RAGTruth train split (MIT): real model outputs with human span annotations, collapsed to response-level. Zero label noise and zero teacher cost. The ~9% synthetic remainder exists for targeted spec-corner coverage — above all abstention, which the spec makes a PASS case and real RAG data barely contains.

2. The synthetic filter gate is a coarse defect catcher, not a quality certificate. The gate was run twice over the same 1,400 candidates:

filter judge	family vs teacher	BAcc	recall(FAIL)	specificity(PASS)	rejected
Claude Opus 5	same	0.67	1.00	0.33	8/1400 (0.6%)
GPT-5.6	cross	0.83	1.00	0.67	57/1400 (4.1%)

(30 human-labelled RAGTruth-test items, identical rubric and call path.)

Inter-judge agreement was 96.5%, but 48 of the 49 disagreements were one-directional (same-family PASS → cross-family FAIL) versus 1 the other way. The same-family judge was lenient toward its own family's prose, and only on the PASS side. The shipped data uses the cross-family gate. Counterweight: GPT-5.6's own specificity is 0.67, so some of those 48 rejections are its false alarms rather than real defects; and judging is batched 10-per-context, so clean responses are seen beside corrupted ones — a contrastive signal absent at inference.

Decontamination

RAGTruth test source_ids excluded at fetch, plus a normalised-response match against every held-out file, which caught 103 real leaks (HaluBench embeds a RAGTruth slice). Paraphrase-level overlap between train and test contexts is not detected by either gate.

Provenance and licensing

Contexts and the real component derive from RAGTruth (MIT). Synthetic responses were generated by Claude Sonnet 5 and Claude Fable 5 over those contexts and filtered by GPT-5.6. Check the current terms of those model providers before redistributing the synthetic subset or training a competing model on it.

Build

Deterministic — hash-ordered, no RNG.

uv run dataset/build_dataset.py real
uv run dataset/build_dataset.py synthetic --judge-backend codex --judge-model gpt-5.6-sol
uv run dataset/build_dataset.py assemble

---

Training dataset — context-adherence guardrail

Private artifact; not for redistribution. Built by dataset/build_dataset.py (deterministic: hash-ordered sampling, no RNG).

Recipe (hybrid, per Label-by-Construction vs Generate-then-Label lean)

• real: RAGTruth train split (MIT), quality=good, human span annotations -> response-level label (FAIL iff any span). Real model outputs (GPT/Llama/Mistral era generators), zero label noise.
• synthetic: teachers {'claude-sonnet-5': 742, 'claude-fable-5': 601} (via claude -p; per-record provenance in meta.generator — the mix is a quota artifact that doubles as the multi-teacher bias mitigation the literature recommends) over RAGTruth train contexts. Taxonomy grid: 6 constructed-FAIL error types (entity/numeric/temporal/negation/fabrication/causal) + 4 PASS types (grounded, paraphrase, abstention, hard-PASS); style alternates per cell; label by construction, all 10 cells generated in one call per context. Filter stack: JSON/format gate -> near-dup gate (SequenceMatcher>0.9 within source) -> blind judge agreement gate: gpt-5.6-sol (eval rubric verbatim, batched per context, shown only (context, response) — never the intended label or grid cell) must reproduce the constructed label.
• decontamination: dropped 103 items whose normalized response text appears in any held-out file (RAGTruth-test leakage via HaluBench etc.); RAGTruth test source_ids excluded at fetch. 2 exact dups removed.
• class balance: 50/50 PASS/FAIL within each task type (subsample majority, hash order). Metric is balanced accuracy; held-out imbalance left intact.
• truncation: none stored; the 24k-char context cap is applied at prompt render time (eval/prompts.py), identically in training and eval.
• slices: nested 250 c 1000 c 4000 c full, stratified by (task, label).

Filter stats (synthetic)

{
  "teacher_models": {
    "claude-sonnet-5": 742,
    "claude-fable-5": 601
  },
  "judge_model": "gpt-5.6-sol",
  "backend": "cli",
  "n_contexts_ok": 140,
  "n_generated": 1400,
  "n_format_reject": 0,
  "n_quota_abort": 0,
  "n_near_dup": 0,
  "n_judge_disagree": 57,
  "n_judge_error": 0,
  "n_kept": 1343,
  "per_cell_kept": {
    "entity_swap": 140,
    "numeric_error": 140,
    "temporal_error": 139,
    "negation": 139,
    "fabricated_detail": 140,
    "causal_logical": 138,
    "grounded": 137,
    "grounded_paraphrase": 110,
    "abstention": 138,
    "hard_pass": 122
  }
}

Slices

slice	n	FAIL	synthetic	QA/Summary/Data2txt
250	250	125	31	86/80/84
1000	1000	501	97	344/316/340
4000	4000	2000	370	1378/1266/1356
10710	10710	5355	1009	3688/3392/3630

Known limits

• Real component shares the RAGTruth domain with the primary held-out set (in-distribution); FaithBench + HaluBench remain the OOD checks.

• Synthetic FAILs are injected, not natural, errors (MiniCheck protocol: train-on-injected, test-on-natural).

• The filter gate was rebuilt mid-project, and the first version was weak. Both numbers are recorded here on purpose. The gate first ran with a same-family judge (Opus 5, the teachers being Claude), which rejected only 8/1400 constructed labels — a number that reads like teacher excellence and is not. Benchmarking both candidate judges against 30 human-labeled RAGTruth-test items (identical rubric and call path) showed why:

  filter judge	family vs teacher	BAcc	recall(FAIL)	specificity(PASS)	rejected
  Opus 5	same	0.67	1.00	0.33	8/1400 (0.6%)
  GPT-5.6 (gpt-5.6-sol)	cross	0.83	1.00	0.67	57/1400 (4.1%)

  Both catch every hallucination; the entire gap is false alarms on grounded text. The shipped gate is the cross-family GPT-5.6 one.

• The judge swap is itself this project's cleanest self-preference measurement. Both judges scored all 1400 identical candidates: they agree 96.5%, but the disagreements are almost perfectly one-directional — 48 of 49 are items the same-family judge passed and the cross-family judge failed, versus 1 the other way. A same-family judge was measurably lenient toward its own family's prose, and only on the PASS side. Rejections concentrate in the two hardest PASS cells (grounded_paraphrase 30, hard_pass 18), which is where drift is most likely — heavy rewording slipping past the context, dense specifics getting one number wrong. Counterweight, stated plainly: GPT-5.6's own specificity is 0.67, so an unknown share of those 48 are its false alarms rather than real defects; dropping them is the conservative choice and costs 4% of the synthetic half.

• Responses are judged 10-at-a-time against one shared context, so clean ones are seen beside corrupted ones — a contrastive signal absent at inference, which inflates agreement for either judge. The gate is a defect catcher, not a quality certificate.

• Constructed PASS responses are cleaner than real model outputs, which subtly drift. This is the exact reason the set is a hybrid: the RAGTruth-train half (human-labeled, real generations) carries the realistic distribution, while the synthetic half exists for targeted error-type and spec-corner coverage (notably abstention, which real data barely contains).

• Teacher and judge are now cross-family as originally intended (Claude teachers / GPT-5.6 judge). Neither ever touches the headline numbers anyway: every held-out set is human-annotated, so no LLM judges the trained model, and labels come from construction rather than from the judge (MiniCheck ships no verification judge at all). The judge only decides what gets dropped.