ArenaMax

the FINAL release: the pareto frontier of the pool as a 7,000-pair tier — token efficiency, visual preference, scalable projects, methodical/elimination reasoning, honesty, long context, deep work, logical interaction — plus the 12k rule-verified math lane, the s1-style diamond, and a 200-prompt shipped eval

ArenaMax is a preference + SFT distillation dataset built from decisive human votes (arena battles, annotator scores) and rule-verified math (reference-answer exact match). Zero model-written labels, zero judge-written labels — every label is human or programmatic. The crowd's known biases (verbosity tilt, calibration confounds) are measured and disclosed, not hidden, and the shipped views and tiers carry the corrected signal.

Configs

config	rows	what it is
preference_v1	85,650 pairs	the full pool (pair format)
preference_tight_v1	28,281	the tight discipline rule (v1.2: English-only gate removed)
preference_balanced_v1	60,843	the verbosity-balanced slice
preference_diamond_v1	1,000	the s1-style ultra-premium cut
preference_pareto_v1	7,000	the pareto frontier tier (FINAL)
sft_v1	85,642	the pool as chat rows — 104,123,867 gpt2 tokens
sft_math_v1	12,000	the rule-verified math lane — 3,460,189 gpt2 tokens
sft_diamond_v1	1,000	diamond replay in chat format
sft_pareto_v1	9,000	pareto replay in chat format

Format

Preference rows (preference_*):

{"id": "...", "prompt": "...", "chosen": "...", "rejected": "...",
 "meta": {"source": "...", "license": "...", "chosen_model": "...",
          "rejected_model": "...", "language": "...", "category": "...",
          "quality_tier": "...", "verbosity": "...", "difficulty": "...",
          "calib": {...}, "facts_chosen": 0, "facts_rejected": 0, ...}}

SFT rows (sft_*): {"id", "messages": [{role, content}...], "gpt2_tokens", "tags", "meta"}.

The meta block carries the full measured record: source, license, both models, language, category, battle date, turn count, code flag, the discipline gates (disc/dense/struct + _pass + _rej per side), fact counts per side, calibration confound, hedge/budget/capability, quality tier, verbosity posture, difficulty bucket, and the pair_id join back to the preference pool.

Label provenance — all human-decisive or rule-verified

source	rows	license	label signal	era
arena25 (lmarena-ai/arena-human-preference-140k)	71,450	cc-by-4.0 (prompts) / provider-TOU (outputs)	human votes, decisive only — ties and both_bad dropped	2025-04-17 .. 2025-07-24 (100% of rows)
repochat	598	cc-by-4.0	human votes, decisive only	2024-11 .. 2025-02
helpsteer3	13,602	cc-by-4.0	human annotator scores, decisive |overall| >= 2 only	2025 annotation
openmath2	12,000	cc-by-4.0	rule-verified: the solution's boxed answer must match the reference (judge-free)	static (GSM8K + MATH lineages)

Admission gates: decisive wins only; refusal winners dropped (regex, 100% of rows); minimum 40 chars on both sides and context turns; dedup via md5 triple + prompt/chosen ids, plus global cross-conversation near-dup families (minhash 8-gram Jaccard >= 0.8) capped at 3 members.

Considered and excluded: arena-expert-5k (upstream serialization defect — numpy array reprs in parquet string columns), search-arena-24k, webdev-arena-preference-10k, lmsys-chat-1m.

The crowd tilt is disclosed, not deleted

Measured on the shipped pool:

• 61.1% of chosen-side wins are calibration-confounded (calib_confound_share) — the crowd often rewards confidence over substance.
• Verbosity tilt: chosen is longer in 38,048 pairs, shorter in 27,321, parity 20,281 (chosen_tok_longer_share 0.567; chosen median 709 tokens vs rejected 571).
• facts_won_share 0.616 — substance wins only modestly more often than it loses.
• Boilerplate openers appear on the chosen side in 6.1% of pairs.

The views (meta filters, no new data) carry the corrected signal:

view	rule	n
clean_signal	meta.calib.confound == False	33,311
strong	meta.quality_tier == 'strong'	13,711
substance_won	facts_chosen >= facts_rejected	52,752
style_won	meta.cov_r_in_c >= 0.8	12,727
hs3_strong	source == helpsteer3 and meta.strength >= 3	4,300
sft_tight	sft rows whose pair is in preference_tight (pair_id join)	28,281

Coverage

• Pair categories: knowledge_reasoning 35,255 / multilingual 31,010 / dev_interaction 15,123 / writing 4,262.
• Languages: 111 detected; code-tagged prompts: python 1,045, js/html/css 683, cpp 158, csharp 144, java 141, c 102, ts 89.
• Difficulty buckets: t1 9,408 / t2 10,659 / t3 65,583.
• Quality tiers: contested 55,722 / single 16,217 / strong 13,711.

The tiers

• tight — the tight discipline rule (discipline + structure gates). In v1.2 the English-only admission gate was REMOVED (the old rule made tight 100% English); every other gate is byte-identical.
• balanced — all shorter-posture pairs + seeded (20261102) equal samples of parity and longer, so the verbosity tilt cannot vote.
• diamond — the s1-style ultra-premium 1,000 (s1 / LIMA / AlpaGasus lineage: difficulty + diversity + quality selection). Selected only from shipped bytes; fully deterministic — QA re-runs the selection and bit-matches every id.
• pareto — the FINAL tier: every selected pair sits on at least one measured frontier axis (token efficiency, visual preference, scalable projects, methodical/elimination reasoning, honesty, long context, deep work, logical interaction) and passes the full pareto core. Seed 20261104; deterministic; rows are byte-identical replays of the shipped bytes.

Eval v2 (train-disjoint)

The 200-prompt eval v1 leaked the training tiers (19/200 inside the pareto tier, 130 inside balanced). eval_prompts_v2.json re-cuts it train-disjoint: 200 arena prompts from rows outside balanced/tight/diamond/pareto and outside eval v1 (strata 82/73/35/10, prompt floor 16 gpt2) + 50 exact-match math prompts from the held-out math_val split. Zero overlap measured against all six tiers; deterministic (QA bit-match, md5 410cfc253dfc).

Fine-tuning evidence (D9) — measured, including the negative

Three-arm chain at 0.5B LoRA (identical hyperparams; eval subset 55 arena + 15 math prompts, train-disjoint vs both arms; rules-only scoring):

• P1 discipline — TRUE: boilerplate + refusal drops from 34.5% (base) to 9.1% (sft_v1 arm) / 7.3% (pareto arm). Fine-tuning transfers discipline.
• P2 frontier — FALSE: the pareto arm does NOT beat the plain-v1 control at teaching scale (tokens 669 > 503, distinct-3 0.703 < 0.722, math 26.7% < 33.3%; only fence compliance wins, 20.0% > 14.5%).
• P3 length — FALSE: both arms inflate (+35.9% / +80.8% vs the +10% gate).

Honest reading: at 0.5B / 2 epochs / 1,200 rows per arm, fine-tuning on ArenaMax transfers discipline strongly, but the pareto tier does not yet beat the v1 control — frontier-teaching stays unproven at this scale. Training survived five environment resets via the hub mirror; AdamW momentum is not checkpointed, so training re-warms after each resume (conservative direction).

Loading

from datasets import load_dataset

pairs = load_dataset("IO-AIR/arenamax", "preference_pareto_v1")["train"]
chats = load_dataset("IO-AIR/arenamax", "sft_v1")["train"]

License

Per-source (see provenance): cc-by-4.0 for repochat, helpsteer3 and openmath2; arena25 prompts are cc-by-4.0 while outputs remain under their providers' terms of use. Because the outputs of one source carry a provider TOU, the repo declares license: other rather than a single blanket claim — the per-row meta.license field always names the exact terms.