Datasets:
Submission: ten Rev-gen1 models (general), 150M–25.6B
Submission: ten Rev-gen1 decision models
Please move the following ten entries to the main leaderboard as general-purpose decision models. Mode: general. These are the exact model names used in our technical report.
We evaluated 10 frozen Rev-gen1 decision models on the official all/test split: 400 cases and all 2,000 decisions. Each invocation receives a state and all five original questions and returns full probability distributions. Evaluation ran on Penn Betty, using one NVIDIA B200 per model.
These general-purpose decision models were trained on a mixed decision library. Checkpoints and calibration were frozen before this evaluation; no fitting was performed on this test evaluation.
| Model | Retained text parameters | Accuracy | KL ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|---|---|
| Rev-gen1-UltraBERT-150M | 151,033,345 | 0.7630 | 0.173438 | 0.091960 | 0.227861 |
| Rev-gen1-EmbeddingGemma-2-Frozen | 272,454,401 | 0.6015 | 0.248282 | 0.142162 | 0.041574 |
| Rev-gen1-UltraBERT-380M | 377,538,561 | 0.7620 | 0.144075 | 0.075490 | 0.202611 |
| Rev-gen1-nano | 752,393,024 | 0.7515 | 0.116630 | 0.063719 | 0.093639 |
| Rev-gen1-micro-lite | 1,312,822,496 | 0.7645 | 0.105980 | 0.058067 | 0.106707 |
| Rev-gen1-micro | 1,881,825,088 | 0.7645 | 0.149561 | 0.080450 | 0.222567 |
| Rev-gen1-mini-lite | 2,759,498,048 | 0.7775 | 0.100136 | 0.055138 | 0.109810 |
| Rev-gen1-mini | 4,205,751,296 | 0.7760 | 0.096074 | 0.052343 | 0.121429 |
| Rev-gen1-midi-lite | 15,736,622,944 | 0.7885 | 0.083206 | 0.046126 | 0.123546 |
| Rev-gen1-midi | 25,624,600,064 | 0.7945 | 0.085678 | 0.046821 | 0.126710 |
Accuracy is agreement with the benchmark gold argmax. KL, Brier, and 10-bin ECE use the pinned upstream scorer. These scores measure agreement with the benchmark teacher distributions.
The seven generative models use the existing 600K checkpoint selection; Rev-gen1-EmbeddingGemma-2-Frozen uses its previously selected epoch-4 head. Probability temperatures are unchanged from the frozen model releases; no checkpoint, temperature, or other setting was fitted on this test run.
Both UltraBERT entries use seed 20261016, explicitly selected before this evaluation: 150M uses cross-lr10 at 600K visits and 380M uses cross-lr10-w5k at 450K visits. The latter is the only 380M seed that passed the original 0.001 probability-invariance gate. These are individual checkpoints, not the paper’s three-seed averages.
Benchmark revision: 94993b820b6d445cec3b2312511904c7eee2af9f.
Measured local inference time
These measurements cover a direct adapter call for one complete five-question case after the qualification warm-up, including tokenization, device transfers, and synchronized GPU execution. They exclude model loading and HTTP/network overhead, so they are not directly comparable with API round-trip latency.
| Model | Median ms/case | 95th percentile ms/case | Mean ms/case |
|---|---|---|---|
| Rev-gen1-UltraBERT-150M | 81.8 | 88.2 | 82.3 |
| Rev-gen1-EmbeddingGemma-2-Frozen | 91.8 | 158.0 | 101.0 |
| Rev-gen1-UltraBERT-380M | 81.7 | 87.8 | 82.2 |
| Rev-gen1-nano | 51.8 | 57.3 | 79.9 |
| Rev-gen1-micro-lite | 36.9 | 44.8 | 66.8 |
| Rev-gen1-micro | 52.1 | 58.2 | 79.7 |
| Rev-gen1-mini-lite | 220.8 | 254.8 | 236.1 |
| Rev-gen1-mini | 361.0 | 415.3 | 374.2 |
| Rev-gen1-midi-lite | 82.2 | 105.5 | 198.9 |
| Rev-gen1-midi | 673.0 | 780.9 | 688.1 |
Execution detail: Rev-gen1-mini, Rev-gen1-midi, Rev-gen1-mini-lite used the unchanged canonical TreeScorer in BF16 after the optimized candidate failed the predeclared training-fixture parity gate. The other generative models used the bucketed scorer after passing that gate. EmbeddingGemma uses its preserved frozen encoder and trained-head adapter; UltraBERT uses the frozen cross-encoder with parallel criterion batches within each question. No tolerance was relaxed. Canonical replay checks establish repeatability, not optimized parity. These execution differences also affect the measured latency.
Availability and verification
These are creator-reported results from preserved research checkpoints; the weights and evaluation code are currently private. This discussion supplies the scores and protocol for the requested leaderboard entries and does not claim public open-weight availability.
All eleven evaluated systems (these ten entries and one separate reference) passed independent local rescoring and result-backup verification. Each submitted model covers every official test case and question; the largest rescoring difference was 2.8e-17. The separate reference is not part of this submission.
The official test Parquet SHA256 is 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c. The current dataset revision 992ce442963b8983deedc94a85048396c5fee633 has byte-identical test data and scorer to the evaluated revision. The scores above were evaluated on 2026-11-02 UTC (2026-11-01 America/New_York).
Machine-readable submitted metrics and prediction hashes
[
{
"model": "Rev-gen1-UltraBERT-150M",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 151033345,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.763,
"kl_from_gold": 0.173437767053537,
"brier": 0.09196006345115426,
"ece": 0.2278612116186287,
"predictions_sha256": "8975c34f41d988f919272237bf7e8664f7f3129fd5965275e53706d14d8699cb"
},
{
"model": "Rev-gen1-EmbeddingGemma-2-Frozen",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 272454401,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.6015,
"kl_from_gold": 0.24828154042932124,
"brier": 0.14216221524664482,
"ece": 0.04157367513310673,
"predictions_sha256": "245535dd660935725ce00569abfe868303766adaa38f22acb4adc204b31672a9"
},
{
"model": "Rev-gen1-UltraBERT-380M",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 377538561,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.762,
"kl_from_gold": 0.14407481842922454,
"brier": 0.07548958342178784,
"ece": 0.20261059255568434,
"predictions_sha256": "8ffbb69dacd411a69f90133f4a5371a479198abff71d68038ab94c226904a79d"
},
{
"model": "Rev-gen1-nano",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 752393024,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7515,
"kl_from_gold": 0.11663005277767577,
"brier": 0.06371866397039364,
"ece": 0.09363862647303306,
"predictions_sha256": "0a71b96dfe37716a58dcd028abdc5b06f47d4252f35d2dde4fc066247dc05cfa"
},
{
"model": "Rev-gen1-micro-lite",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 1312822496,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7645,
"kl_from_gold": 0.10598042724233314,
"brier": 0.058066748395719764,
"ece": 0.10670726066395433,
"predictions_sha256": "d4f67a406bd84d2dfeb7c11684940d6be97cf46021b7fc634e277cd7d54190e3"
},
{
"model": "Rev-gen1-micro",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 1881825088,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7645,
"kl_from_gold": 0.14956052582175147,
"brier": 0.08045044308864865,
"ece": 0.22256733442563098,
"predictions_sha256": "54d6325a1cb4939723af860d7042e449355f04b6e704595e28c64322d743ce69"
},
{
"model": "Rev-gen1-mini-lite",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 2759498048,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7775,
"kl_from_gold": 0.1001362968334701,
"brier": 0.0551380328405775,
"ece": 0.10981009193557663,
"predictions_sha256": "a89e304e321b9fd9cbac5f99579ffbe04159977cd947454f4897d2ac9075128f"
},
{
"model": "Rev-gen1-mini",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 4205751296,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.776,
"kl_from_gold": 0.09607390843090029,
"brier": 0.052343351562307176,
"ece": 0.12142904791869148,
"predictions_sha256": "8cdd7b1a6f1d5e00484b6ddb9fb37bdd837564019cfeba3c819c11961dd58f87"
},
{
"model": "Rev-gen1-midi-lite",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 15736622944,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7885,
"kl_from_gold": 0.08320633483386591,
"brier": 0.04612558816167158,
"ece": 0.12354594034763633,
"predictions_sha256": "33f8cfef407fe36cf00326e25999558f55fce2f6c3a4da354a5230ce75cc2a0f"
},
{
"model": "Rev-gen1-midi",
"mode": "general",
"requested_leaderboard": "main",
"parameters": 25624600064,
"n_cases": 400,
"n_decisions": 2000,
"accuracy": 0.7945,
"kl_from_gold": 0.08567779051968133,
"brier": 0.04682061223831304,
"ece": 0.12670998931055483,
"predictions_sha256": "a3d698248908230c00f7e602f8edeb0a0731f3274edaeff314bf98555eec6860"
}
]
Thanks for the detailed protocol. All ten Rev-gen1 models are now in the fitted table (PR #65), from Rev-gen1-midi at #13 (0.7945) down to Rev-gen1-EmbeddingGemma-2-Frozen at #43 (0.6015). They're marked as trained on a mix that includes filtered train, with latency shown as your local B200 timings.
Since the weights and predictions are private, we couldn't rescore them, and the card says so for these rows. If you publish the prediction files later (the hashes you listed would let anyone check them), we'll rescore with bench/score_with_jobs.py and drop that note.
The canonical-state audit against test is a nice touch; mid-size models like micro-lite at 37 ms per case sitting within about 0.03 of the 25.6B midi is the interesting result here.
@sandeep85 The earlier fitted-on-train request was in error. The split we used for these checkpoints did not use the training data. Please see my recent request to move it to the other table. Thanks.