chaoliangUNSW/MacJev-0.8B-Decision-Data

MacJev-0.8B decision training data (round 2, 2026-09-24)

This repository holds the data used to train MacJev-0.8B. MacJev-0.8B is a Qwen3.5-0.8B decision model. For each option it outputs a score, read as h[slot] · (w_yes − w_no) at a -> slot placed after that option. It was trained on inputs of up to 25,600 tokens.

The repository also holds the evaluation results of the one-H100 training run.

Layout

Path	Contents
pool/qwen35/	The assembled training pool used for the run (Qwen3.5 tokenizer). {train,dev,cal}.rows.jsonl hold the records; *.ids.npy hold the pre-rendered token ids. manifest.json, epoch_plan.json and assembly_report.json describe how the pool was built.
pool/minicpm5/	The same pool rendered for MiniCPM5 (the fallback arm of the base bake-off; not used by the final model).
pool/ledger.jsonl	One line per candidate row: kept or dropped, with the reason.
raw/<category>/	Per-category builder outputs (train/dev/cal) and manifests. Categories: typed_official, typed_synthetic, long_typed, general_*, intent_*, nli, theme_* (spam, phishing, jailbreak, toxicity, relevance, triage, routing), long_retrieval, long_qa, long_table, mac_gate, mac_goal_done, mac_legacy.
teacher_jobs/	Teacher-model jobs and outputs: typed labels (Claude Opus 5.5 / Sonnet 5, GPT-5.6-Sol), typed workflow designs, NLI translations, theme data, zh filler corpus, Mac goal paraphrases.
deliveries/typed_synth_burst_2026-09-24/	Additional synthetic typed decisions (train/dev only).
exclusions/	Hashes used to keep evaluation rows out of training.
eval_results/	Results of the round-2 H100 run (see below).

Row schema: see pool/qwen35/manifest.json. Every record has id, group_id, split, source, category, lang, license, state, question (t = choice / score / noul, ins, crit), valid and, where a teacher distribution exists, soft.

What was left out

• All held-out test sets, including our own sealed suites: long_grid, long_grid_plus, mac_gate_test, mac_goal_done_test, typed_synth_unseen and typed_long_unseen.
• Every teacher job line and delivery row belonging to the 9 held-out typed workflows.
• Public-dataset test.jsonl splits.

Keeping these out means the suites stay usable for future evaluation.

Sources and licences

Each record carries its own license field. Rows come from:

• public datasets: MASSIVE, CLINC150, HWU64, MNLI, AG News, DAIR Emotion, GoEmotions, SST-5, BoolQ, SQuAD v2, HotpotQA, QuALITY, civil_comments, Enron spam, phishing emails, several jailbreak and prompt-injection sets, and LocalLLaMA/typed-decisions;
• teacher-generated data (Claude Opus 5.5, Claude Sonnet 5, GPT-5.6-Sol);
• programmatically generated data (Mac simulator states, synthetic tables, long-context retrieval grids).

Some upstream licences are unclear; they are marked unclear: in the license field. Check the upstream terms before reusing a subset.

Evaluation results (eval_results/)

• h100_run_20260924/: pipeline state, configs, logs and training curves for every run (base bake-off, LR pilot, main, held-out ablation). It also has calibration temperatures and row-level predictions on the public sealed test suites (typed-decisions test, AG News, DAIR Emotion, banking77, MASSIVE-51, XNLI, bench_apps themes, option-order flips).
• laya_scoreboard/: paired comparison with Laya's three official checkpoints on identical rows, with 95% case-cluster bootstrap CIs.
• own_suites/: aggregate metrics on our own sealed suites (long_grid / 25K claim, Mac gate, unseen typed workflows). Only aggregates are included, no rows.
• external/: JevBench v1.4.1 public-item results.

Headline numbers (main model, sealed tests):

Metric	MacJev-0.8B	best Laya	Jev
typed-decisions accuracy (2,000)	0.791	0.766	0.727
typed-decisions Brier vs soft labels (lower is better)	0.046	0.061	0.148
banking77, 77 labels (never trained)	0.682	0.492	—
MASSIVE 51-language intent macro	0.717	0.401	—
MASSIVE, 37 never-trained languages	0.655	0.361	—
guardrails jailbreak (bench_apps)	0.905	0.805	—
moderation toxicity (bench_apps)	0.725	0.535	—
49-suite macro ECE, as deployed (lower is better)	0.054	0.299	—
JevBench v1.4.1 public accuracy (231 items)	0.641	0.584	0.866
tweet_topic, zero-shot	0.755	0.632	0.793
fin_topic, zero-shot	0.467	0.342	0.670
daily_dialog, zero-shot	0.325	0.614	0.710

• Jev's typed-decisions number comes from the typed-decisions dataset card.
• Jev and Laya numbers for JevBench come from the v1.4.1 results file.
• Jev and Laya numbers for the zero-shot sets come from the elcronos/jev-vs-open-decision-models study.
• Where MacJev is weaker than Laya (AG News, phishing, RAG relevance, support triage, held-out Emotion and SST-5), the full table in laya_scoreboard/scoreboard.md says so.

25,000-token claim. Both the preregistered long_grid and the amended long_grid_plus pass. At 24K the controlled accuracy is 0.553, against 0.564 in the 2K–4K reference bins. All 10 depth deciles are within 10 pp of the reference, and all controls are at chance. The amended suite has 128 items per decile; it was amended before any model was scored on it.