Datasets:
Benchmark results for 17 trending decision models (zero-shot + 2 fitted, HF Jobs run 2026-10-16)
We scored the trending decision-model models on Typed Decisions following the card's protocol, and would like to submit the rows below for the leaderboard.
Protocol: test split, 400 cases / 2,000 decisions, one request per case with the state and all five questions, full distributions scored. All runs executed on Hugging Face Jobs on 2026-10-16. Raw per-case predictions, per-workflow/per-question breakdowns and the aggregated leaderboard are public in the results bucket (leaderboard.md, results/<model>.json).
Metric conventions (same scorer for every row, including our Prior reference): accuracy = argmax agreement (gold ties count as correct); KL(gold ‖ pred), predictions clipped at 1e-9; per-decision multi-class Brier; ECE over 10 equal-width confidence bins. Reference rows with this scorer: Prior 0.4865 / KL 0.327 / Brier 0.181 / ECE 0.042; Uniform accuracy 0.308.
General models, zero-shot
| Model | Kind | Accuracy | KL ↓ | Brier ↓ | ECE ↓ | PR |
|---|---|---|---|---|---|---|
| perplexity-ai/pplx-decider-v1.1-27b | general, zero-shot | 0.753 | 0.441 | 0.160 | 0.060 | #5 |
| vllm-sr/d3 (27B) | general, zero-shot | 0.749 | 0.409 | 0.146 | 0.049 | #4 |
| vllm-sr/d3-mini (4B) | general, zero-shot | 0.743 | 0.319 | 0.123 | 0.047 | #5 |
| Cloudflare/clef (27B) | general, zero-shot | 0.728 | 0.191 | 0.103 | 0.025 | #12 |
| openjev/openjev (27B) | general, zero-shot | 0.728 | 0.506 | 0.191 | 0.101 | #6 |
| vllm-sr/d3-flash (9B) | general, zero-shot | 0.717 | 0.322 | 0.131 | 0.031 | #2 |
| Cloudflare/clef-flash (9B) | general, zero-shot | 0.710 | 0.209 | 0.111 | 0.016 | #7 |
| autotrust/JEV-35B (35B MoE) | general, zero-shot | 0.700 | 0.526 | 0.169 | 0.062 | #1 |
| autotrust/JEV-27B-VL (28B) | general, zero-shot | 0.688 | 0.544 | 0.194 | 0.113 | #9 |
| Cloudflare/clef-omni (35B) | general, zero-shot | 0.673 | 0.247 | 0.123 | 0.030 | #3 |
| LiquidAI/d1-3B | general, zero-shot | 0.656 | 0.286 | 0.144 | 0.037 | #11 |
| vllm-sr/d3-nano (2B) | general, zero-shot | 0.629 | 0.545 | 0.211 | 0.113 | #1 |
| vllm-sr/d3-lite (0.8B) | general, zero-shot | 0.531 | 0.518 | 0.244 | 0.130 | #2 |
| autotrust/GEV-26B-Decide (26B) | general, zero-shot † | 0.522 | 0.423 | 0.215 | 0.066 | #3 |
| autotrust/JEV-27B (27B) | general, zero-shot † | 0.495 | 0.550 | 0.253 | 0.136 | #2 |
| vllm-sr/d3-edge (0.6B) | general, zero-shot | 0.425 | 1.176 | 0.456 | 0.341 | #2 |
| doofz/THX-01 (0.3B) | general, zero-shot | 0.377 | 0.620 | 0.320 | 0.203 | #1 |
† GEV-26B-Decide and JEV-27B were scored through their native POST /v1/decide interface, which is one decision per request (the card's own note on request shape applies: this is not the one-request-per-case shape). Option/rubric descriptions were embedded in the question text, since their APIs take options as plain strings. All other rows used the single-request /v1/systemone shape (vLLM nightly).
Fitted or fine-tuned on train (second table)
| Model | Kind | Accuracy | KL ↓ | Brier ↓ | ECE ↓ | PR |
|---|---|---|---|---|---|---|
| mrm8488/embeddinggemma-2-ft-typed-decision (0.3B) | fine-tuned on train |
0.770 | 0.119 | 0.065 | 0.155 | #3 |
| nandakishorm/vega-08b-public-intents (0.8B) | specialist (bundled typed-decisions adapter active, gate 1.0) | 0.663 | 0.435 | 0.191 | 0.074 | #1 |
Serving notes: vLLM-served rows used vLLM nightly /v1/systemone with --max-model-len 8192 --max-num-seqs 8; THX-01 via thx01.decide(), vega via vegaml (mode engine), embeddinggemma via its model.decide() API. Latency figures were measured on heterogeneous Jobs hardware and are not comparable across rows, so we omit them.
Happy to rescore with a different input mapping if any author prefers.
Reproduction commands — all runner scripts are public in the results bucket under scripts/ (score_model.py, setup_vllm.sh, setup_py.sh), so any run reduces to one hf jobs run:
1 — vLLM-served models (/v1/systemone, one request per case — the card's protocol):
hf jobs run --flavor rtx-pro-6000 --timeout 2h --detach --secrets HF_TOKEN \
--volume hf://buckets/SaylorTwift/typed-decisions-evals:/mnt/bucket \
--env BENCH_MODE=vllm --env BENCH_MODEL=vllm-sr/d3 --env BENCH_NAME=vllm-sr_d3 \
--env BENCH_OUT=/mnt/bucket/results \
--env "BENCH_SERVE_FLAGS=--max-model-len 8192 --max-num-seqs 8 --logprobs-mode raw_logprobs" \
pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel \
bash /mnt/bucket/scripts/setup_vllm.sh
setup_vllm.sh installs vLLM nightly (--extra-index-url https://wheels.vllm.ai/nightly) and runs score_model.py, which starts vllm serve $BENCH_MODEL, polls /health, then POSTs all 400 test cases to /v1/systemone sequentially and scores the full distributions. Smaller models used the same command with --flavor swapped (l4x1, l40sx1, a100-large, h200 as needed).
2 — Package-native models (same script, setup_py.sh):
# THX-01
--env BENCH_MODE=thx01 --env BENCH_MODEL=doofz/THX-01 --env "BENCH_EXTRAS=thx01"
# vega 0.8B (noul mapped to vega's "boolean" type, mode "engine")
--env BENCH_MODE=vegaml --env BENCH_MODEL=nandakishorm/vega-08b-public-intents --env "BENCH_EXTRAS=vegaml torch transformers"
# embeddinggemma-2 typed-decision
--env BENCH_MODE=eg2td --env BENCH_MODEL=mrm8488/embeddinggemma-2-ft-typed-decision --env "BENCH_EXTRAS=torch transformers pillow sentencepiece"
3 — Per-question native interfaces (GEV-26B, JEV-27B):
# GEV-26B: documented transformers+peft System 1 path (one forward pass per question)
--env BENCH_MODE=gev --env BENCH_MODEL=autotrust/GEV-26B-Decide --env "BENCH_EXTRAS=torch transformers peft safetensors accelerate" # a100-large
# JEV-27B: repo's serve_decide.py (POST /v1/decide), one request per question
--env BENCH_MODE=jevdecide --env BENCH_MODEL=autotrust/JEV-27B # rtx-pro-6000, setup_vllm.sh
4 — Prior reference row (no model):
hf jobs run --flavor cpu-basic --env BENCH_MODE=prior --env BENCH_OUT=/mnt/bucket/results \
--volume hf://buckets/SaylorTwift/typed-decisions-evals:/mnt/bucket \
python:3.12-slim-bookworm bash /mnt/bucket/scripts/setup_py.sh
5 — Aggregate into leaderboard.md (pulls results/*.json from the bucket, renders both tables, uploads):
python aggregate_results.py results-download
Each job writes results/<BENCH_NAME>.json directly into the bucket via the mounted volume — containing the full request protocol metadata, per-case predictions, per-decision scores (accuracy/KL/Brier/ECE per question), and aggregate metrics. Metric code and the Prior validation live in score_model.py; the scorer loads load_dataset("LocalLLaMA/typed-decisions", "all", split="test") and sends {"state": ..., "questions": ...} verbatim from the row, so no prompt engineering is involved for /v1/systemone models.
Update: vllm-sr/d3-turbo (26B, just released) scored with the same protocol and scorer — added to the bucket leaderboard:
| Model | Kind | Accuracy | KL ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|---|---|
| vllm-sr/d3-turbo (26B) | general, zero-shot | 0.736 | 0.438 | 0.162 | 0.069 |
Notes: right at the teacher self-agreement ceiling (0.735). Served through the model's native AutoModel.system_one() runtime (transformers), because the current vLLM nightly does not yet support its decision config (attention_mode 'prefix_bidir' — vLLM serve exits at startup). One request per case, same as all other rows. Eval-results PR: vllm-sr/d3-turbo (.eval_results).