n1h111sm/add-rm-criteria

What this is

Attribute criteria extracted for every preference pair in n1h111sm/add-rm-data, for the text-based-optimization extension to ADD-RM (Attribute-aware Domain-adaptive Distributional Reward Model).

Each row is a source preference pair plus 2-4 short natural-language criteria naming the abstract qualities that explain why the chosen response was preferred.

⚠️ These are ORACLE criteria — read before using

The criteria were generated with the preference label visible: the extractor saw the query, both responses, and which one a human preferred, then named the attributes that explained the choice.

Valid uses

• Initialisation for a criterion set in an attribute-structured reward model
• Distillation target for a label-free criteria predictor
• Analysing which attributes drive preference in each domain

Invalid use

• Any evaluation that conditions on these criteria is contaminated. Accuracy gains measured this way reflect label leakage, not method quality. They upper-bound what criteria-conditioning could achieve; they do not measure a deployable system.

A contamination audit on a 200-pair sample found content-word overlap with the chosen response of 0.035 vs 0.011 with the rejected one (asymmetry +0.024). So the criteria are genuinely abstract rather than fingerprints of the chosen text — but they remain label-conditioned.

Schema

column	type	notes
idx	int64	row position in the source split; joins back to add-rm-data
domain	string	e.g. shp-askengineers_train
prompt	list[{role, content}]	1 user turn
chosen	list[{role, content}]	[user, assistant], preferred
rejected	list[{role, content}]	[user, assistant], dispreferred
criteria	list[string]	2-4 abstract criteria, <=8 words each

Splits

split	rows	domains
shp	348,671	18

Generation

• Extractor: Qwen/Qwen3-4B-Instruct-2507 (natively non-thinking) via vLLM, greedy (temperature=0, max_tokens=128).
• Prompted for abstract qualities only — no proper nouns, no facts or phrases from either response, no reference to "Response A/B", max 8 words per criterion.
• Post-filtered to drop A/B references, cap length, and de-duplicate within an example.

Known characteristics

• 348,669 of 348,671 rows have exactly 4 criteria.
• ~405k distinct criterion strings, but with heavy surface-form redundancy — "concrete actionable steps", "provides concrete actionable steps", and "actionable steps" are separate strings for the same concept. Semantic canonicalisation is required before treating the frequency table as a concept distribution.
• The head is concentrated: technical depth (45,897) and acknowledges uncertainty (29,340) together cover ~22% of examples.
• The source SHP data carries a strong length confound — P(chosen longer) ~ 0.675 — so criteria mentioning detail or thoroughness may be tracking length rather than quality.

Update — criteria_query_only column

A second criteria column generated from the question alone.

column	generator sees	label-free?
criteria	query + both responses + which was preferred	✗ oracle
criteria_query_only	query only	✓

criteria_query_only is the proper label-free analogue of an attribute router q(w | phi(x)): the extractor is shown only the question and must predict which qualities would separate an excellent answer from a poor one, having seen no answers.

Same model (Qwen3-4B-Instruct-2507), same greedy decoding, same output format and length caps as the oracle column. The only difference is the information available, so any downstream gap between the two columns is attributable to information rather than to prompt style or generator capability.

Unlike criteria, this column carries no label leakage and is valid for evaluation.

Update — extractor-size columns (*_30b)

Both criteria columns regenerated with a 7.5x larger extractor, to test whether a more capable criteria producer closes the label-free gap.

column	extractor	generator sees	label-free?
criteria	Qwen3-4B-Instruct-2507	query + both responses + which was preferred	no (oracle)
criteria_query_only	Qwen3-4B-Instruct-2507	query only	yes
criteria_30b	Qwen3-30B-A3B-Instruct-2507-FP8	query + both responses + which was preferred	no (oracle)
criteria_query_only_30b	Qwen3-30B-A3B-Instruct-2507-FP8	query only	yes

The two extractors share the same Instruct-2507 post-training recipe, chat template (natively non-thinking), prompt text, greedy decoding, and output-format caps. Only parameter count differs, so column-to-column differences isolate extractor scale.

The criteria* oracle columns are label-conditioned: valid as a ceiling, as initialisation, or as a distillation target, but invalid for any evaluation that conditions on them. The *_query_only* columns carry no leakage.