Nutrition Preference Comparisons

Summary and intended purpose

Pairs and small sets of candidate responses to the same clinical nutrition task, annotated with the criteria under which one candidate is preferred. Intended to support preference optimisation workflows and to reveal systematic model behaviours.

Target training or evaluation tasks

• Pairwise preference between candidates on a stated rubric
• Identifying the candidate that correctly acknowledges missing information
• Preferring constraint-compliant over superficially fluent responses

Clinical scope

• Adult clinical nutrition tasks mirroring the instruction dataset
• Comparison criteria focus on constraint adherence, completeness, and safety-relevant omissions

Example record schema

{
  "example_id": "npc-0001-synthetic",
  "specialty": "clinical_nutrition",
  "task_type": "pairwise_preference",
  "language": "en",
  "scenario": {
    "summary": "Adult meal-structure task with a stated energy target and a stated food intolerance.",
    "stated_constraints": [
      "energy_target_range",
      "intolerance_exclusion"
    ]
  },
  "candidates": {
    "a": {
      "response": "<candidate_a>",
      "constraints_satisfied": [
        "energy_target_range"
      ]
    },
    "b": {
      "response": "<candidate_b>",
      "constraints_satisfied": [
        "energy_target_range",
        "intolerance_exclusion"
      ]
    }
  },
  "preference": {
    "chosen": "b",
    "criteria": [
      "constraint_adherence",
      "completeness"
    ],
    "confidence": "high",
    "label_source": "rubric"
  },
  "review_status": "automated_only",
  "generation_version": "gen-v2"
}

Generation methodology

1. 01Multiple candidates are generated per task under varied conditions.
2. 02Rubric-based scoring produces a preference label and the criteria that drove it.
3. 03Ties and low-confidence comparisons are retained with explicit flags rather than discarded.

Validation information

• Preference labels in the research corpus are rubric-derived; they are not individually expert-assigned.
• Expert-labelled subsets can be scoped for engagements that require them.

Known limitations

• Not every preference label comes from a clinical expert.
• Rubric-based preferences may reward rubric compliance rather than clinical quality.
• Commercial availability and licence terms must be scoped before use.

Licensing and access process

Licence status: Research access by inquiry. Access is sales-assisted. Licensing is handled by agreement, based on data provenance, rights, permitted uses, delivery, and commercial terms.

1. Step 1Inquiry describing the post-training setup and target behaviours.
2. Step 2Discussion of rubric alignment and whether expert-labelled subsets are needed.
3. Step 3Rights review and written access terms.