vllm-sr/decision-2.0-decision-index

Decision 2.0: Decision Index 0.2.1 runs

Complete runs of the released Decision 2.0 models on the Decision Index 0.2.1 suite, scored with the unmodified kit at 87d4650 (decision_index score --edition 0.2.1).

Model	Revision	Decision Index	Raw index	Breadth skill	Requests ok	unsupported
Decision 2.0 Vega 27B	7aec49ae	56.47	66.91	55.47	150,317	0
Decision 2.0 Lux 9B	78bf3c03	46.26	59.02	44.95	150,315	2
Decision 2.0 Nox 4B	25e8f67d	43.77	57.24	42.10	150,315	2
Decision 2.0 Sol 2B	64235bef	29.53	45.94	26.77	150,315	2
Decision 2.0 Eos 0.8B	3594047d	20.15	39.04	17.55	150,315	2
Decision 2.0 Kai 0.6B	cd49ea38	16.29	36.05	14.24	150,299	18

Every run is complete: scores.json says "complete": true, with a result for each of the 150,317 scoreable requests. Every request is ok except the few a model declines because the prompt is longer than its input limit (unsupported, reason max_length_exceeded; they count as wrong). No request errored or abstained. Nothing was truncated and no options were removed.

How the runs were made

• Engine. In-process, publication.decision_index_release_engine:ReleasedDecisionIndexEngine (a copy is in harness/decision_index_release_engine.py). It passes each request's original state and questions to the package's own Decision2.system_one, the same entry point as AutoModel.from_pretrained(..., trust_remote_code=True).system_one on the model cards. One request per call, all questions of a request in one batch, raw probabilities (temperature 1, no calibration). A question over the package's input limit makes the request unsupported; a malformed answer would be an error (there were none). The kit's runner is unmodified and was run with --compact.
• Hardware and software. One AMD Instinct MI325X (256 GB) per runner process, one request at a time. ROCm 7.2, PyTorch 2.12.0, Transformers 5.17.0, flash-linear-attention 0.5.2, causal-conv1d 1.7.0; BF16 backbone and FP32 decision head.
• Two passes per model, one package. The 38 index benchmarks (120,226 requests) were run first, split over 2 to 8 runners. The 30,091 requests of the six benchmarks the board shows but does not count (MMLU, ARC-Easy, ARC-Challenge, SimpleBench, RouterBench and SGD) were run afterwards by one more runner (8 for Vega 27B), with the same engine, package, container, kit and frozen kernel-autotune cache. harness/receipt.json has the row accounting: every scoreable run ID exactly once, from one model source. harness/runners/ has every runner's environment.json and status.json, with rows paths shortened to file names. The 442 requests the kit excludes at scoring time were not run.
• Weights. Each run's package has the same weights fingerprint (identity.model_sha256 in MODEL_MANIFEST.json) as the revision listed above; see harness/weights-vs-release.json. The packages are the frozen release candidates, so their environment.json names the pre-release repository and its revision. Later revisions changed only the runtime code, the card and the repository name. Each was published only after its answers matched the previous revision's on our release panels: the latest runtime update matched on 10,653 of 10,653 prompts.

Spot check of the released revisions

Each spot-checked revision was downloaded from the Hub (every file checked against its manifest) and run with the same engine, container and kit on a stratified sample of 2,160 scoreable requests (the first 50 of each benchmark in SHA-256 order of their run IDs; SimpleBench has 10). Its answers were compared with the stored rows: a request agrees when its status and every question's chosen option are equal. Results per model are in harness/release-spotcheck.json. No status differed. Every disagreement is a single question whose top two options were within 0.025 of each other in one of the two runs (listed with both margins under flipped_questions), the size of the numerical differences between runtime revisions.

Model	Revision	Requests that agree	Largest probability difference
Decision 2.0 Vega 27B	7aec49ae	2,158 / 2,160	1.0e-02
Decision 2.0 Lux 9B	78bf3c03	2,159 / 2,160	1.1e-02
Decision 2.0 Nox 4B	25e8f67d	2,159 / 2,160	1.4e-02
Decision 2.0 Sol 2B	64235bef	2,158 / 2,160	1.4e-02
Decision 2.0 Eos 0.8B	3594047d	2,157 / 2,160	1.2e-02
Decision 2.0 Kai 0.6B	cd49ea38	2,160 / 2,160	0

Nox, Lux and Vega are complete runs on byte-identical weights (harness/weights-vs-release.json), scored with the pre-release runtime, like the other three. The first runtime revisions with the fast kernel path hit a ROCm Triton compile error (PassManager::run failed) on long requests with many questions, which then returned error: Nox ce1bdc9d, Lux 214ffa43 and Vega 9b067a95 on 5 requests of this sample (9.7K tokens and longer), and Eos ad0aa724 and Sol 4b75b521 on one full-Index request outside it. The cause was a kernel that chose between two tensors at run time when only one of them was larger than 2 GiB. The runtime-only revisions with the fix (Vega 477e90f5, Lux 7c6792f7, Nox 7fc0023a, Sol 23cbe9f9, Eos 34e2db97) have the same weights, and their runtime also falls back to the plain PyTorch path, which gives the same values, if a fused kernel ever fails to compile. The revisions listed here (Vega 7aec49ae, Lux 78bf3c03, Nox 25e8f67d, Sol 64235bef, Eos 3594047d, Kai cd49ea38) are the next runtime-only revisions, again with the same weights. Their runtime keeps every captured GPU graph instead of evicting one when its graph cache is full: on ROCm, evicting large graphs under long, varied traffic could crash the GPU process (a stress run over this index's requests crashed the previous Eos runtime on 6 of 6 shards; the new runtime finished the 2 shards it ran). They answered our release panels exactly as the previous revisions did (10,653 of 10,653 prompts, also with the graph cache cut to 8 graphs), pass the spot check above with no status change, and answer the 60 heaviest Index requests as the previous revisions did, without an error.

Files (runs/<model>/)

• results.jsonl.gz: one row per request in the kit's format, without payload. Every row keeps its payload_sha256. The runner ran with --compact, so rows never held payload or raw_output.
• scores.json, index.json, benchmark-summary.json, score.log: decision_index score --edition 0.2.1 on results.jsonl.gz.
• harness/: receipt.json (row accounting and digests), complement-panel.json (what the second pass ran), weights-vs-release.json, release-spotcheck.json (the spot check of the listed revision), and runners/<runner>/environment.json and status.json.

Re-scoring

Build the 0.2.1 suite with the kit (suite rebuild, then suite import; see its README), then:

python -m decision_index score --edition 0.2.1 --engine decision-2.0-vega-27b \
    --results runs/decision-2.0-vega-27b/results.jsonl.gz --out rescored/decision-2.0-vega-27b

This reproduces scores.json, index.json and benchmark-summary.json, apart from generated_utc.

Running a model with the kit

pip install "transformers>=5.17" torch safetensors   # plus flash-linear-attention and causal-conv1d for the kernel path
hf download vllm-sr/Decision-2.0-Nox-4B --revision 25e8f67d1b486c647222df3aac640d2d5d736bbe --local-dir nox
export DECISION2_PACKAGE_DIR=$PWD/nox
PYTHONPATH=harness python -m decision_index run --engine decision_index_release_engine:ReleasedDecisionIndexEngine \
    --option model_id=vllm-sr/Decision-2.0-Nox-4B --option revision=25e8f67d1b486c647222df3aac640d2d5d736bbe \
    --option package_manifest_sha256=$(sha256sum nox/MODEL_MANIFEST.json | cut -c1-64) --option device=cuda:0 \
    --rows <rows.jsonl.gz> --out runs/nox --compact

model_id must equal the package's repo_id. Vega 27B also needs the pinned Qwen/Qwen3.8-27B base, from the Hugging Face cache or DECISION2_BASE_DIR. A Hub download answers as Decision-2.0-<name> in response.model; the scored release candidates answered under their pre-release names (DEV2.0-<size>).

License

The run files are Apache-2.0, like the models. They hold no benchmark text: each row has run IDs, answers, probabilities and timings only.