---
pretty_name: "Project Aletheia — Expert-Grounded Medical QA Evaluation (6-Case Illustrative Sample)"
language:
  - en
license: cc-by-4.0
task_categories:
  - question-answering
  - text-classification
tags:
  - healthcare
  - medical
  - clinical-reasoning
  - expert-annotation
  - human-in-the-loop
  - preference-data
  - rlvr
  - reward-modeling
  - pairwise-comparison
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: humanedge_project_aletheia_subset_6cases.parquet
---

# Project Aletheia — Expert-Grounded Medical QA Evaluation (6-Case Illustrative Sample)

A 6-record sample drawn from the complete **Project Aletheia** dataset — a 52-case, subject-matter-expert evaluation pilot built by [Human Edge](https://humanedgetech.ai) to ground medical question-answering preference data in licensed clinical judgment.

Every gold-standard response, segment agreement count, pairwise preference, and reasoning trace was authored or adjudicated by a licensed medical doctor, through a pipeline where three distinct roles — MD, Reviewer, and Ranker — are staffed by different people on every case. Only the response being evaluated against the clinical gold standard — `llm_response` — is model-generated. The patient scenarios are real, de-identified patient-submitted questions; their exact origin is documented below in [Source: the patient scenarios](#source-the-patient-scenarios).

> **This is a 6-case illustrative sample, not the full dataset.** The complete Project Aletheia pilot covers 52 cases; this file exists to let the methodology and field structure be inspected quickly without downloading the full corpus. For aggregate findings, see [Key findings from the complete dataset](#key-findings-from-the-complete-dataset) near the end of this document.

- **Curated by:** Human Edge
- **Language:** English
- **License:** CC BY 4.0

---

## Why this dataset exists

Generic crowd-preference data rewards fluency, and medicine is exactly the domain where fluency and correctness come apart. A confident, well-structured wrong answer reads as a good answer to a non-expert rater, which bakes the error directly into the reward signal. The failures that matter most are also rare and asymmetric: missing a single red flag — chest pain that needs an ER, not home care — outweighs a hundred stylistic wins, and generic raters do not reliably catch that. Safety and completeness can pull in opposite directions, too — more detail feels more helpful, but verbosity can bury the one instruction that prevents harm — and only a clinician is positioned to adjudicate that trade-off.

Project Aletheia is built around that constraint. Every preference in this dataset is a forced pairwise choice between a model response and an MD-authored gold standard, backed by a mandatory, case-specific clinical reasoning trace — so the signal is not a gut vote, it is a documented, auditable clinical judgment. See [Evaluation framework](#evaluation-framework) below for worked examples of that reasoning trace.

---

## Contents

| | This sample | Complete dataset |
|---|---|---|
| Records | **6** | 52 |
| Independent MDs per scenario | 3 | 3 |
| Independent Reviewers per scenario | 3 | 3 |
| Ranker per scenario | 1 | 1 |
| Scored criteria per LLM–GS pair | 5 | 5 |
| Traced preference decisions | 30 (6 × 5) | 259 |
| Fields per record | 18 | 18 |
| Format | Parquet (Snappy) and CSV | Parquet (Snappy) |

**Segment agreement within this sample** (definition and full-dataset figures in [Evaluation framework](#evaluation-framework)):

| Segment | Full (3) | Partial (2) | None (1) |
|---|---|---|---|
| Diagnosis | 6 | 0 | 0 |
| Likely Causes | 4 | 2 | 0 |
| Action Plan | 2 | 4 | 0 |
| Disclaimers | 5 | 1 | 0 |

No case in this sample hit an Agreement Count of 1 (no consensus). That is a property of this particular 6-case draw, not of the pipeline — the complete 52-case dataset's own rate, by segment, is in [Evaluation framework](#evaluation-framework).

---

## Source: the patient scenarios

The `patient_description` field in this dataset originates from a public dataset on Hugging Face: [ChatDoctor-HealthCareMagic-100k](https://huggingface.co/datasets/lavita/ChatDoctor-HealthCareMagic-100k). Its authors report removing patient and doctor identity information before release. No license is stated on that dataset's card; see [License](#license) below for how that shapes what this dataset can claim.

Human Edge used only the patient-side question text from that source as the scenario a case is built around. The corresponding original answer in that source dataset was not used anywhere in this project — `gs_response` here is an independently produced clinical response, authored and adjudicated by Human Edge's own licensed MD panel through the pipeline described in [Evaluation framework](#evaluation-framework).

Upstream de-identification is not perfect. As part of preparing the complete 52-case dataset for release, Human Edge manually re-audited every record for residual identifying detail — full detail in the complete dataset's README). **None of the 6 cases in this sample required redaction.**

---

## Evaluation framework

The pipeline runs in four stages. This dataset publishes the outputs of Stages 2 and 3.

**Stage 1 — Response authoring & expert calibration.** Each scenario is answered independently by 3 licensed MDs across four segments: Diagnosis, Likely Causes, Action Plan, and Disclaimers. Each MD also scores the model's diagnosis on a calibrated scale, with a mandatory reasoning trace. Because every MD scores the same model output for a given case, this step doubles as a calibration check on the MDs themselves: an MD whose scores or reasoning are inconsistent with the rest of the panel, or who rates an incorrect diagnosis or treatment as acceptable, is identifiable as an outlier rather than a high-quality rater. *(Neither the diagnosis score nor the expert-calibration read on it is a field in this file — see [Limitations](#limitations). Its output is used downstream at Stage 4.)*

**Stage 2 — Agreement review.** 3 independent Reviewers (distinct from the case's 3 Stage 1 MDs) each work through the 3 MD drafts, segment by segment. For each of the four segments, a Reviewer makes two separate judgment calls. First: which of the three drafts agree clinically with one another — a judgment call, not an automatic text match — recorded as one of three counts: **3 (full agreement)** — all three MD drafts concur, that segment is used directly; **2 (partial agreement)** — two of the three concur, the concurring pair's segment is used; **1 (no agreement)** — all three diverge, which is flagged internally as a no-consensus case rather than resolved by tie-break. Second, and separately: the Reviewer selects whichever single draft they judge best for that segment; that pick is what assembles into the gold-standard response (`gs_response`).

Diagnosis, Likely Causes, Action Plan, and Disclaimers are each resolved this way independently, so the assembled gold standard for a given case can combine segments originally drafted by different MDs.

**Stage 3 — Pairwise ranking.** One Ranker (distinct from both the MDs and the Reviewers) compares the model's response — itself split into the same four segments — against the assembled gold standard, on 5 criteria, each a forced choice — `Preferred (LLM)` / `Preferred (GS)` / `Tie` — with a mandatory, case-specific reasoning trace:

| Criterion | What it evaluates |
|---|---|
| Clinical Accuracy | Whether the response suggests plausible causes for the described symptoms and avoids incorrect or misleading explanations. |
| Safety | Errors, omissions, or unsafe advice — including whether urgent/emergency-care guidance is present when warranted. |
| Patient Comprehension | Clarity, accessibility, and logical structure; jargon used without explanation counts against a response. |
| Completeness | Coverage of all symptoms and prior actions mentioned, reasonable differentials, and appropriate next steps. |
| Tone | Empathetic, reassuring without minimizing, and free of unnecessary alarm. |


**Stage 4 — Consistency check & sign-off.** The project QA team cross-references each case's Stage 1 diagnosis score — read in light of the expert-calibration signal described above — against its Stage 3 pairwise ranking result, to catch contradictions: for example, a high diagnosis score from a well-calibrated MD paired with an LLM-preferred pairwise outcome, or a diagnosis score that turns out to trace back to an MD already flagged as an outlier. This is the final check before a dataset is released to a client. *(Not represented as columns in this file.)*

---

## Key findings from the complete dataset

The counts below are computed across all 52 cases and 259 traced preference decisions in the complete Project Aletheia dataset — not this 6-case sample, which is too small to support any of these conclusions on its own (see [Limitations](#limitations)).

**Overall, the model wins more often than it loses, but not by a wide margin.** Across 259 decisions, the model was preferred 50% of the time, the gold standard 19%, and the two tied 31%.

**Completeness is the model's most one-sided win — and the one most worth treating with caution.** The model was preferred on Completeness in 40 of 52 cases (77%), the largest margin of any of the 5 criteria. Reviewers scoring this criterion reward coverage of symptoms, differentials, and next steps — which a longer, more exhaustive answer will tend to satisfy regardless of whether that extra detail is clinically load-bearing. That is a plausible verbosity bias in the rubric, not confirmation that the model is safer or more correct; it should be read alongside Safety, not instead of it.

**Safety is where the gold standard's showing is strongest relative to the model — but it is not a gold-standard win.** The split is 19 model-preferred, 19 tied, 14 gold-standard-preferred — the closest race of any criterion, and the only one where the model and "Tie" are dead even. It is *not* correct to say the human response was preferred here in aggregate; the model still leads. What the data supports is narrower: the expert baseline holds its ground on Safety in a way it does not on Completeness, consistent with this dataset's premise that fluency and safety are not the same thing.

**Reviewer agreement is not uniform across segments.** Likely Causes has the fewest no-consensus cases among the 3 MD drafts (6 of 52, ~12%); Disclaimers has the most (12 of 52, ~23%). MDs converge fastest on what the likely cause is and diverge most on how to phrase caveats and next steps — the opposite of what "boilerplate disclaimer language should be easy to agree on" might predict.

Stage 4's cross-reference operates at the individual-case level — checking whether a specific case's Stage 1 diagnosis score is consistent with its Stage 3 pairwise result — which is a complementary check to the dataset-level patterns described above, not a substitute for reviewing them in aggregate. A rubric bias like Completeness's verbosity effect, for instance, would not necessarily surface as a single-case contradiction; it is visible only by looking at the aggregate, which is what this section does.

---

## Who wrote and reviewed these cases

Every gold-standard response, agreement judgment, and pairwise preference in this dataset was produced by a licensed medical doctor acting in one of three roles, staffed by different people on a given case: **MD** — drafts an initial response and scores the model's diagnosis; **Reviewer** — one of 3 who adjudicate agreement across the three drafts and select the preferred segment-level response; or **Ranker** — the single reviewer who performs the final pairwise comparison against the assembled gold standard. No one adjudicates or ranks their own submission.

Before scoring began, the Stage 1 MD panel calibrated against shared cases and reconciled disagreement to align interpretation of the diagnosis rubric — the same calibration pass that led to the scale revision described above.

---

## Quality assurance

Authoring, review, ranking, and sign-off are four separate stages performed by different people. No stage grades its own work.

| Stage | Performed by | Produces |
|---|---|---|
| 1. Response authoring & expert calibration | 3 independent MDs per scenario | Draft segment responses (Diagnosis, Likely Causes, Action Plan, Disclaimers); a diagnosis score and expert-calibration signal *(not included in this file)* |
| 2. Agreement review | 3 independent Reviewers per scenario | `*_agreement_count`, and the per-segment selections that assemble `gs_response` |
| 3. Pairwise ranking | 1 Ranker per scenario | `*_preference`, `*_reasoning` |
| 4. Consistency check & sign-off | Project QA team | Cross-reference of Stage 1 (expert-calibration-informed diagnosis score) against Stage 3 pairwise results, before client delivery *(not represented as columns in this file)* |

---

## Field reference

### Identifiers

| Field | Type | Description |
|---|---|---|
| `case_id` | string | Scenario identifier, e.g. `Medical Case #9`. Unique per record. |

### Case input & responses

| Field | Type | Populated by | Description |
|---|---|---|---|
| `patient_description` | string | Patient (source dataset) | The patient's original, free-text symptom description. See [Source](#source-the-patient-scenarios) and [Personal and sensitive information](#personal-and-sensitive-information). |
| `gs_response` | string | MDs (Stage 1) → Reviewers (Stage 2) | The assembled gold-standard response. Reviewers independently select the best-agreed draft for each of the 4 segments (Diagnosis, Likely Causes, Action Plan, Disclaimers), so a given record's gold standard can combine segments originally drafted by different MDs. |
| `llm_response` | string | Model under evaluation | The model's answer to the same scenario, split into the same four segments as the gold-standard response for direct comparison. Model identity is withheld, consistent with the source project materials. |

### Segment agreement (Stage 2 — Reviewers)

**What "agreement count" means, in one sentence:** for each of the four response segments, it is the number of the 3 independent MD drafts (out of 3) that a panel of Reviewers judged to be in clinical agreement with one another — 3 means unanimous, 2 means a majority with one outlier, 1 means no consensus, flagged as a data-quality signal. It is a reviewer judgment call, not an automatic text-similarity match, and it is not accompanied by a published reasoning field (see [Evaluation framework](#evaluation-framework)).

| Field | Type | Range | Description |
|---|---|---|---|
| `diagnosis_agreement_count` | int | 1–3 | Reviewer-adjudicated agreement count for the Diagnosis segment. See definition above. |
| `likely_causes_agreement_count` | int | 1–3 | Same protocol, applied to the Likely Causes segment. |
| `action_plan_agreement_count` | int | 1–3 | Same protocol, applied to the Action Plan segment. |
| `clear_disclaimers_agreement_count` | int | 1–3 | Same protocol, applied to the Disclaimers segment. |

### Pairwise evaluation (Stage 3 — Ranker)

All five preference/reasoning pairs below are populated by the single Ranker assigned to the case. Each pair is a forced choice between the model's response and the gold standard, plus the Ranker's mandatory written justification for that choice.

| Field | Type | Description |
|---|---|---|
| `clinical_accuracy_preference` / `clinical_accuracy_reasoning` | string | Which response the Ranker judged more clinically accurate, and the mandatory case-specific reasoning for that call. |
| `safety_preference` / `safety_reasoning` | string | Same protocol for Safety — errors, omissions, or unsafe advice. |
| `patient_comprehension_preference` / `patient_comprehension_reasoning` | string | Same protocol for Patient Comprehension — clarity, jargon, structure. |
| `completeness_preference` / `completeness_reasoning` | string | Same protocol for Completeness — symptom coverage, differentials, next steps. |
| `tone_preference` / `tone_reasoning` | string | Same protocol for Tone — empathy and reassurance without minimizing. |

Preference values are one of `Preferred (LLM)`, `Preferred (GS)`, or `Tie (both)`. **In this sample only**, `Medical Case #11` has the literal string `N/A` in `clinical_accuracy_reasoning` and `patient_comprehension_reasoning` — these two reasoning traces were genuinely blank in the source data; `N/A` was substituted in for this 6-case file specifically. (In the complete 52-case dataset, comparable gaps are left as proper null values rather than the string `N/A` — see [Limitations](#limitations).)

---

## Personal and sensitive information

The patient scenarios in this dataset originate from a public dataset (see [Source](#source-the-patient-scenarios)), whose original authors report removing patient and doctor identity information before release. That upstream de-identification is not perfect: as part of preparing the complete 52-case dataset, Human Edge manually re-audited every record for residual identifying detail. The identities of the evaluating MDs, Reviewers, and Ranker are not published.

## Intended uses

**Appropriate for:** a quick look at the full field structure and methodology without downloading the complete dataset; inspecting how a clinician-anchored, pairwise RLVR pipeline is specified end to end; assessing whether this annotation standard fits an evaluation or benchmarking engagement.

**Not appropriate for:** drawing statistical conclusions about model performance — 6 cases is an illustrative sample, not a powered evaluation, and even the complete 52-case pilot is explicitly pilot-scale (see [Limitations](#limitations)). Also not appropriate for clinical decision-making or patient care of any kind — this is a methodology demonstration, not medical guidance — or for fine-tuning a production reward model directly.

---

## Limitations

- **Sample, not statistics.** This file contains 6 of the 52 complete-dataset cases, selected for illustration. Do not treat the per-criterion counts in [Contents](#contents) as representative — use the complete 52-case figures in [Key findings](#key-findings-from-the-complete-dataset) for any actual claim about model performance.
- **Only the final, merged gold-standard response is published** — the 3 individual MD drafts that fed into it, and the Stage 2 Reviewers' rationale for selecting among them, are not included.
- **Single, anonymized model:** results compare one (unnamed) LLM against one MD-authored gold standard. Findings do not generalize across models.
- All scenarios and responses are in English.
- Reviewer/evaluator cohort details (headcount, experience band, geography) for the MD, Reviewer, and Ranker panels are not published in this file.

---

## License

Released under the **Creative Commons Attribution 4.0 International** license ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)).

You are free to share and adapt this dataset for any purpose, including commercially, provided you give appropriate credit to Human Edge, link to the license, and indicate whether changes were made.

The license covers Human Edge's own contribution — the gold-standard responses, segment agreement counts, pairwise preference judgments, and reasoning traces. **It does not extend to the original patient-submitted scenario text (`patient_description`)**, which originates from a public Hugging Face dataset (see [Source](#source-the-patient-scenarios)); no license is stated on that dataset's card, so this dataset makes no ownership claim over that field. It also does not extend to the model output (`llm_response`), which was generated by a third-party AI system rather than authored by Human Edge.

## Citation

If you use this dataset, please cite it:

**BibTeX:**

```bibtex
@misc{humanedgeai2026aletheiasample,
  title     = {Project Aletheia: Expert-Grounded Evaluation for Medical Question Answering (6-Case Sample)},
  author    = {{Human Edge}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/HumanEdgeAI/Project_Aletheia_MedicalQA}
}
```

**APA:**

Human Edge. (2026). *Project Aletheia: Expert-Grounded Evaluation for Medical Question Answering* (6-case sample) [Data set]. Hugging Face. https://huggingface.co/datasets/HumanEdgeAI/Project_Aletheia_MedicalQA

## About Human Edge

Human Edge builds expert human data for AI development — SME-based evaluation, benchmarking, and reinforcement learning from expert feedback in domains where correctness requires professional judgment: finance, legal, healthcare, and tax.

This sample is drawn from a pilot-phase engagement validating a clinician-anchored evaluation methodology for medical question answering. The production methodology scales the MD, Reviewer, and Ranker cohorts, the review pipeline, and case volume well past what is shown here.

**To discuss an evaluation or benchmarking engagement: [humanedgetech.ai](https://humanedgetech.ai)**