teilomillet/dmbench-examples

DMBench examples

90 structured examples: nine per category, drawn from the rebuilt typed-decisions-fr-v4 development benchmark. Each of the ten categories has three Choice, three Score and three Noul examples, spanning short, medium and long prompts within that category and response type.

These are clear fictional French service and workplace scenarios. Each row contains the complete model input, answer options or rubric, the reference answer, a concise explanation, a native reference response and its evaluation.

• Read the worked examples, or browse the examples split above.
• See three Score reranking comparisons: laptop returns, live ticket availability and catering offers. Each uses a selected Score row and describes an alternative candidate under the same user demand and rubric.
• File hashes and the source identity are in MANIFEST.json.

Reference responses are authored, gold-derived illustrations. No evaluated model produced them. Their one-hot probabilities explain the response format; they are not measured confidence or evidence of model accuracy. Their passing evaluations follow by construction. There are no model results or rankings here.

What the categories mean

Category	Practical decision
conditions	Check eligibility against explicit service requirements.
action	Choose a useful next step, respect prerequisites and avoid duplicate effects.
quand-appeler	Decide whether an external tool is needed, authorized and ready.
securite	Check identity, authority and operation scope; assess established incident status.
routage	Match a request to the responsible human service.
preuves	Assess support from relevant, current, competent and validated sources.
contraintes	Prioritize mandatory requirements over optional preferences.
mise-a-jour	Follow the current request after corrections, additions or cancellations.
donnees	Check verified recipients and permission for each disclosed field.
escalade	Choose a human handoff and its operational priority.

The supplied rules define each exercise. They do not claim to describe universal law or real incidents, and no external legal or medical knowledge is required.

Load the data

import json
from datasets import load_dataset

examples = load_dataset("teilomillet/dmbench-examples", split="examples")
row = examples[0]

# Send only this object to the model.
request = json.loads(row["request_json"])
print(json.dumps(request, ensure_ascii=False, indent=2))

# These fields belong to the evaluator.
print(row["expected_answer"])
print(row["explanation"])
print(json.loads(row["reference_response_json"]))
print(json.loads(row["evaluation_json"]))

The same 90 records are available in examples.jsonl. state, instructions and options are display columns. request_json preserves the exact native input and option order: Choice/Noul criteria are mappings, whereas Score criteria are ordered lists. JSON text columns keep those different structures compatible with a single table schema.

IDs, length bands, reference answers, explanations and evaluation fields must stay outside model features. The release IDs do not expose source endpoint orientation, full source family IDs or an unreleased answer pattern.

Choice, Score and Noul

The native wire format is wyat-answer-v1. The name identifies the response contract; it does not attribute these reference responses to Wyat.

• Choice selects one supplied option and provides probabilities and a ranking. Accuracy checks the declared selection against the reference answer.
• Score rates an ordered rubric. It returns a distribution over levels and expected_ordinal_value = sum(index * probability). This build uses three, four or five levels where their distinctions make sense. For example, 20% probability at level 2 and 80% at level 3 produces an expected index of 2.8. Modal-level accuracy, expected-index error, normalized MAE and ranked probability score are separate metrics. A high value can mean suitability, severity or urgency according to the stated axis; raw values across axes are not a common utility.
• Noul evaluates a direct true/false proposition with probability_true. The example evaluator uses >= 0.5 to select true. Noul is not abstention; unknown facts must not silently become false facts.

Score can support reranking: hold the user demand, common context and rubric fixed, score each candidate independently, then sort by the expected score in that axis. The worked comparisons show reference rubric levels; actual model expectations may be fractional. These are illustrative two-candidate comparisons, not listwise retrieval measurements or calibrated cross-query relevance. Documentary support does not by itself measure answer completeness.

Missing, invalid, unsupported and non-OK responses remain failures in evaluation denominators. A null per-example metric means it does not apply to that response type. This preview does not estimate calibration, robustness or general ability.

Selection and release scope

There are exactly three rows per category/response-type panel. Prompt length is the Unicode character count of the serialized native request_json.

For each panel, choose a short prompt and a long prompt from distinct scenario families, then a middle prompt from a third family. Except for three curated Score anchors, these are the shortest, longest and available prompt nearest their length midpoint. The middle Score examples for action, quand-appeler and contraintes retain the laptop-return, ticket-availability and catering cases reviewed before this release; their families are reserved before selecting the short and long examples. The underlying inputs are unchanged.

This is a length-stratified illustration with reviewed anchors, not a random or statistically representative sample. Length ranges are modest because the rebuilt scenarios are deliberately concise. The 90 structured rows use distinct source families. The reranking page additionally describes three alternative candidate summaries from the approved comparisons.

The full 1,200-question suite, complete scenario families, source authoring files, generators, local audit archives and model results are outside this release.

Checks and limits

The source candidate received label-blind LLM readings of all current inputs and pairwise reviews of its scenarios. Ambiguities were corrected and changed inputs were reread. These are authoring checks, not human certification or independent statistical confirmation. The source still has limited scenario coverage and residual surface cues; this small release is not a model-quality evaluation.

The packaged reference responses are parsed and graded with the native scorer. The JSONL schema is checked through Hugging Face datasets/Arrow. Source, scorer and published file hashes are recorded in the manifest. None of these mechanics establishes practical model reliability or training-overlap freedom.

These examples are intentionally exposed. Keep them out of untouched confirmation claims and disclose their use during model development. No model was rerun to produce this release.

Original authored examples and documentation are distributed under the included MIT license. Project: DMBench.