evaluation/README.md

Decision evaluation: 512 questions

The model card compares NVFP4 and Q8_0 Dory with hosted models on these questions. NVFP4 validation results use a separate 1,024-question set.

Download the questions and targets. This is the exact file used for the base Dory, fine-tuned Dory, and hosted-model comparisons in the model card.

Task	Questions	Source and target
State tracking	192	SCONE Alchemy and Tangrams: questions about recorded world states after instructions.
Planning	192	Generated block-stacking and package-delivery problems: simulated legality and search-derived optimal actions.
Human judgements	128	HelpSteer2: questions with distributions derived from released human ratings and preference votes.

These examples were withheld from this model's training. Their task families are represented in training. The results measure performance on this task mix, not general decision-making ability or the source datasets' original benchmarks. Related examples share a group_id; 512 questions should not be interpreted as 512 independent source situations. Hosted providers' training overlap with these sources is unknown.

File integrity

SHA-256 of questions-512.json:

e2ca7413a306fc2b0e9e044aa685c94d5aa8e0def7cfcec6448941b7369ae625

The file is a JSON array. Original identifiers and metadata are preserved so its hash matches the evaluation reports.

Fields

Field	Meaning
id	Unique question identifier.
state	Context presented to the model.
instructions	Question or judgement requested.
kind	noul for binary, choice for named options, or score for ordered levels.
options	Ordered list of [answer_name, answer_description] pairs. Preserve this order.
target	Target probability for each option, in the same order. Never send these values to the model.
target_basis	How the reference distribution was constructed.
source, dataset, domain	Source and task identifiers.
group_id	Related source episode, prompt, or canonical planning problem.
tokens, labels	Dory prompt token IDs and corresponding answer-label token IDs used in local evaluation. Other models use their own tokenizers.
role, component	Dataset assembly metadata; not model input.

Requests

Send one question at a time with its unchanged state, instructions, and answer descriptions. For choice and binary questions, map answer names to descriptions. For scores, send the descriptions as an ordered list. For example, construct an OpenRouter native decision request from a row as follows:

criteria = (
    [description for name, description in row["options"]]
    if row["kind"] == "score"
    else dict(row["options"])
)
request = {
    "model": model_id,
    "state": row["state"],
    "questions": {
        row["id"]: {
            "type": row["kind"],
            "instructions": row["instructions"],
            "criteria": criteria,
        }
    },
}

Hosted comparisons use /api/alpha/decisions at OpenRouter. The llama.cpp comparison uses /v1/systemone. Providers control their internal prompts and probability settings. Local Dory evaluations use two recurrent loops and no fitted calibration. No generated reasoning or additional examples are requested.

Scoring

Read probabilities in options order. For a binary response p, use p for true and 1-p for false. All reported comparisons require a valid response for every question; no failed questions are omitted.

• Accuracy: choose the option with the largest predicted probability. Resolve prediction ties by the original option order. Count it as correct if its target probability is within 1e-6 of the largest target probability.
• Brier error: sum (prediction[i] - target[i])**2 over options, then average over questions. Lower is better. This is a sum over options, not an average over options.
• Combined results: average over all 512 questions, rather than averaging the three task scores equally.

Hosted responses are checked for finite probabilities between zero and one and the exact expected options. If their sum differs from one by at most 0.02, divide by the sum before scoring. Larger discrepancies are rejected. This handles provider rounding without changing the ordering of answers. The hosted comparison report records normalization counts, observed discrepancies, and exact served model versions.

Targets can represent a recorded fact, several equally optimal actions, or disagreement among human raters. Matching their most probable answer does not by itself establish calibrated confidence.

Sources, changes, and licences

The evaluation data have separate terms from the model weights' OpenMDW licence.

• SCONE: Long, Pasupat, and Liang (2016), Sequential CONtext-dependent Execution, under CC BY-SA 4.0. Episodes and observed states were converted into typed questions and reference distributions. These adapted rows remain under CC BY-SA 4.0.
• HelpSteer2: NVIDIA and Scale AI, HelpSteer2, under CC BY 4.0. Prompts, responses, ratings, and preference votes were converted into typed judgement questions and empirical distributions. Source revision: 990b2711a36180dd19d9c94b8627844866f8982a.
• Planning: independently generated block-stacking and delivery problems, inspired by the task families in PlanBench. Published PlanBench instances and generator code were not used. The generated questions and targets are provided by Ryan Daum under CC BY 4.0.

Collection and question-format adaptations are provided under CC BY-SA 4.0; the source-specific terms above remain applicable. Source provenance includes source archive hashes and the split policy.

Results

• Base versus fine-tuned Dory, NVFP4 / Eider
• Base versus fine-tuned Dory, Q8_0 / llama.cpp
• Hosted decision models on the same questions