SOURCE_REVIEW.md

Source review and mixture rationale

This is the source-selection record for the local v1 build. It is deliberately separate from the normalized data so the model artifact can be released without publishing the derived training set.

Included

Source	Role in v1	Important caveat
open-r1/OpenR1-Math-220k	Main math source; multiple reasoning traces and Math-Verify correctness metadata	The builder uses only a bounded streamed subset and creates evidence views; it does not claim every trace is human-verified.
aladinDJ/ultramix-DPO-annotated	Balanced general, reasoning, math, coding and instruction-following preferences with curation metadata	Preference pairs are kept as pairwise supervision, not treated as absolute truth.
nvidia/HelpSteer3	Human-annotated multi-turn preference coverage across general, STEM, code and multilingual domains	Overall preference has a three-point direction scale; ties are skipped.
coseal/CodeUltraFeedback_binarized	Code-specific preference and rating signal	Identical chosen/rejected pairs are skipped.
Skywork/Skywork-Reward-Preference-80K-v0.2	Additional general reward-preference diversity	The source has limited task metadata, so rows use the generic rubric and preserve the upstream source tag.
R-I-S-E/RISE-Judge-SFT-20K	Judge-style comparisons with a generated rationale and position-swap verification	The builder extracts the user question and A/B candidates, discarding the long judge rationale from the scored candidate.
iknow-lab/JudgeBias-DPO-RefFree	Bounded anti-shortcut/materials-science pairs for representation invariance and injected-error detection	The builder uses only the stable validation_1000.parquet file because the full upstream parquet set currently has inconsistent schemas.
Anthropic/hh-rlhf	Multi-turn helpfulness and harmlessness preferences from human feedback	The default Hub config exposes transcript strings, so the builder extracts the shared dialogue prefix and final assistant response. The source contains sensitive content; keep the derived data private unless separately cleared.
stanfordnlp/SHP	Naturally occurring human-written preferences across 18 question/instruction domains	Labels are collective Reddit preferences, not factual correctness. The builder filters weak score-ratio pairs and records subreddit metadata. The card discusses Reddit API terms rather than granting a simple redistribution license.
AIPlans/PKU-SafeRLHF-RLHF	Safety-aware preference pairs with a derived margin for confidence weighting	CC-BY-NC-4.0 is inherited from PKU SafeRLHF; this is suitable for the private research build, not automatically for commercial redistribution.
SUSTech-NLP/MixReward	Multilingual preference coverage across 103 languages and six broad domains	Apache-2.0 card, but the mixture aggregates multiple upstream sources; retain the language and source-subset metadata when auditing or redistributing.
project-themis/Themis-CodePreference	Multi-criterion code preference signal spanning correctness, efficiency, security and maintainability across eight languages	Apache-2.0 card. Aspect and language IDs are mapped into the row instruction and metadata; the source itself is already a curated mixture, so the quota stays bounded to avoid code dominance.

Explored but not in the default build

Source	Decision
OpenAI PRM800K	Deliberately not used as the main math source because v1 is intended to be richer than simple step labels.
Math-Shepherd	Useful process-supervision reference, but not included in the default mixture; its automated step labels would add another label regime before the core schema is stable.
Full JudgeBias train/validation parquet set	Not used because the Hub parquet files currently expose inconsistent columns and fail the streaming cast check; only the stable 999-row validation_1000.parquet slice is included.
amazon/CodePrefBench	Kept as an evaluation candidate rather than training data; its Hub card identifies a non-commercial license and an evaluation/test orientation.
cracklinoatbran/reward_hacking_policy_1073	Prompt-only reward-hacking evaluation set. It needs fresh candidate generations and verified hack/legitimate labels, so it is reserved for a later anti-shortcut generation stage.
hlyn-labs/prompt-injection-judge-dataset-v1	Security-judge conversations are not directly compatible with the v1 candidate-level correctness contract; reserve for a later adversarial/instruction-injection slice.

Mixture properties of the generated artifact

The expanded build contains 53,955 unique candidate rows in three group-isolated Parquet splits:

• 16,000 absolute math-verification rows.
• 37,955 pairwise preference/judge rows across general chat, multilingual, safety, STEM, code, materials science, and math/judgment sources.
• 41,959 rows with no evidence, 3,976 with a reference solution, 4,014 with privileged context, and 4,006 with both.
• 48,607 train, 2,697 validation, and 2,651 test rows.
• 53,955 unique example_id values and 27,219 unique group_id values.
• 12 source datasets, with 160,198,172 bytes across the three Parquet shards.

The source datasets are not all licensed identically. This folder is therefore a private local artifact by default; check each upstream dataset's current license and terms before redistributing derived data. The model-weight release should document the source mixture without publishing these Parquet files unless redistribution is separately cleared.