Rewrite dataset card for the curated train/test split

.gitattributes

@@ -58,30 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
-data/normalized/ctgov_normalized.jsonl filter=lfs diff=lfs merge=lfs -text
-data/raw_metadata/ctgov.jsonl filter=lfs diff=lfs merge=lfs -text
-data/review_queue/review_queue.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/sft_examples.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/sft_examples.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/test.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/test.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/train.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/train.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/validation.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/validation.jsonl filter=lfs diff=lfs merge=lfs -text
-data/normalized/normalized.jsonl filter=lfs diff=lfs merge=lfs -text
-data/raw_metadata/aea.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/sft_curated.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/sft_curated.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/test.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/test.jsonl filter=lfs diff=lfs merge=lfs -text
 data/sft/curated/train.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/train.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/validation.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/validation.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_1of3.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_1of3.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_2of3.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_2of3.jsonl filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_3of3.csv filter=lfs diff=lfs merge=lfs -text
-data/sft/curated/shards/part_3of3.jsonl filter=lfs diff=lfs merge=lfs -text
+data/sft/curated/test.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,78 +1,208 @@
 ---
+pretty_name: HypoAgent — Statistical Analysis Plan SFT Corpus
 license: other
 license_name: mixed-ctgov-public-domain-plus-review
-pretty_name: HypoAgent — Hypothesis-Testing Plan SFT Corpus
 language:
 - en
+task_categories:
+- text-generation
+size_categories:
+- 10K<n<100K
 tags:
 - statistics
 - biostatistics
+- statistical-analysis-plan
 - clinical-trials
 - economics
 - study-design
 - hypothesis-testing
 - instruction-tuning
-task_categories:
-- text-generation
-size_categories:
-- 10K<n<100K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/sft/curated/train.csv
+  - split: test
+    path: data/sft/curated/test.csv
 ---
 
-# HypoAgent — Hypothesis-Testing Plan SFT Corpus (private)
-
-> **🤖 Model trained on this dataset:**
-> [`AliKhajegiliM/HypoAgent-Mixtral-8x7B-Instruct-SFT`](https://huggingface.co/AliKhajegiliM/HypoAgent-Mixtral-8x7B-Instruct-SFT)
-> — a LoRA adapter on Mixtral-8x7B-Instruct-v0.1 (trained on the curated 60k subset). Code &
-> pipeline: private GitHub `AliKhajegiliM/HypoAgent`.
-
-
-> **⭐ For training, use `data/sft/curated/`** — a quality-filtered, diversity-balanced subset
-> capped at **60,000** examples (mean quality 60.1, 0 leakage, 0 method/outcome contradictions).
-> Source-rebalanced to **CT.gov 66% / AEA 25% / OSF 8%** (non-clinical 21%→34%); single-arm capped
-> 24%→14%; 13 designs / 5 outcomes / 5 tasks. Rubric: hard eligibility (outcome determinable, min
-> length, bucket-adaptive quality floor) → keep_score ranking → budgeted allocation (source quotas,
-> per-family cap, 20% per-design cap). The full corpus below is the superset.
-
-
-Supervised fine-tuning corpus of **STUDY CONTEXT → REASONED HYPOTHESIS-TESTING PLAN**
-examples. Each assistant response is *synthesized from extracted structured facts* by a
-transparent statistical-reasoning engine (method keyed on outcome distribution × dependence
-structure × unit × estimand × comparison), not copied from source text. Code + reproducible
-pipeline: private GitHub `AliKhajegiliM/HypoAgent`.
-
-## Key figures
-- **96,026** SFT examples from **27,052** study families; splits **77,076 / 9,316 / 9,634**.
-- **Three sources**: ClinicalTrials.gov **75,499** (public domain) + AEA RCT Registry **15,317**
-  (prospective economics/social-science field & lab experiments) + OSF Registrations **5,210**
-  (non-clinical preregistrations + online A/B). AEA & OSF have unclear content licenses and are
-  tagged `training_use_status=review_required`; their plans are synthesized (no verbatim text).
-- Group split by `study_family_id` → **0** train/val/test leakage.
-- Task types: full plan (A), method selection (B), completion (C), critique of a flawed plan (D),
-  alternatives (H).
-- **0** method/outcome contradictions across all flagship plans (audited).
-- Designs: parallel RCT, single-arm-vs-benchmark, observational, cluster-randomized, crossover,
-  factorial, **field/lab/randomized experiment (AEA)**, online_experiment (OSF).
-
-## Sources & compliance
-CT.gov examples are public domain. AEA (economics field/lab experiments; JEL-coded) and OSF
-(psychology/ecology/social + online A/B) add NON-clinical, prospective diversity but have unset/
-unclear content licenses, so they are tagged `training_use_status=review_required` /
-`human_review_status=license_review`. All assistant plans are **synthesized** from structured
-facts — no verbatim source text. Filter on `metadata.source_platform` and
-`metadata.training_use_status` to include/exclude by source and license posture.
-
-## Record schema (JSONL / CSV)
-`{messages:[system,user,assistant], metadata:{example_id, split, task_type, domain,
-study_design, outcome_type, difficulty, quality_score, study_family_id, source_platform,
-license, training_use_status, method_family, source_ids, source_urls, challenge_sets, ...}}`.
-The CSV flattens the three turns into `system`/`user`/`assistant` columns (list fields joined
-with `|`; read with a real CSV parser — text contains commas/newlines).
+# HypoAgent-SFT: study description → prespecified statistical analysis plan
+
+54,163 supervised fine-tuning pairs that teach a model to read a study registration and write the analysis plan that belongs to it. The input is a study description — objective, design, population, exposure, comparator, outcome, timing, whatever the registry record actually contains. The target is a prespecified hypothesis-testing and statistical-analysis plan built for *that* design.
+
+This is the curated split used to train **[HypoAgent/HypoAgent-Mixtral-8x7B-Instruct-SFT](https://huggingface.co/HypoAgent/HypoAgent-Mixtral-8x7B-Instruct-SFT)**, which on the held-out `test` split below beats base Mixtral-8x7B-Instruct in 93.6% of judged comparisons and cuts plans containing a critical statistical error from 33.8% to 14.4%.
+
+![Corpus composition](assets/corpus_composition.svg)
 
 ## Contents
+
+```
+data/sft/curated/train.csv    48,218 rows · 18,699 study families
+data/sft/curated/test.csv      5,945 rows ·  2,315 study families
 ```
-data/sft/            train/validation/test + sft_examples (union), each as .csv and .jsonl
-data/normalized/     one canonical schema record per study
-data/raw_metadata/   ctgov.jsonl, aea.jsonl, osf.jsonl, europepmc.jsonl  (raw provenance)
-data/review_queue/   items routed to human review
-reports/             distribution + leakage + coverage reports (CSV + HTML)
+
+Two files, nothing else. The split is **grouped by `study_family_id`**: every task derived from one registry entry lands in the same split, so train/test family overlap is exactly **0**. A family yields between one and five examples depending on how much the source record supported.
+
+```python
+from datasets import load_dataset
+ds = load_dataset("HypoAgent/HypoAgent-SFT")
+print(ds)  # train: 48218, test: 5945
 ```
+
+The text columns contain commas, quotes and newlines — read them with a real CSV parser, never by splitting on commas.
+
+## Record schema
+
+| Column | What it holds |
+|---|---|
+| `system` | Fixed system prompt, identical across the corpus |
+| `user` | The study description plus a one-line task instruction |
+| `assistant` | The target plan |
+| `example_id` | `<source>_<registry id>::<task_type>`, e.g. `ctgov_NCT05144607::A_full_plan` |
+| `study_family_id` | `ctgov:` / `aea:` / `osf:` + registry id — **the grouping key for the split** |
+| `split` | `train` or `test` |
+| `task_type` | One of the five tasks below |
+| `study_design` | 13 values, from `parallel_rct` to `online_experiment` |
+| `outcome_type` | `continuous`, `binary`, `ordinal`, `count`, `time_to_event` |
+| `method_family` | Method class the target commits to (`linear model`, `survival`, `ordinal`, …) |
+| `domain` | 13 values; `clinical` 66.2%, `economics / social science` 25.4%, the rest behavioural and social science |
+| `difficulty` | `complex` (69.4%) or `standard` |
+| `quality_score` | 0–100 curation score. Mean 60.1, range 40–95; nothing below 40 was kept |
+| `challenge_sets` | Pipe-separated evaluation slices this example belongs to |
+| `injected_flaw` | The specific error planted in a `D_critique` prompt; empty for every other task |
+| `license` | Source license posture |
+| `human_review_status` | `auto_pending` (CT.gov) or `license_review` (AEA, OSF) |
+| `synthetic_or_extracted` | Always `synthesized_from_extracted_structured_facts` |
+| `source_ids`, `source_urls` | One registry id and one URL per row, pointing back to the original record |
+
+## The five tasks
+
+| `task_type` | n | What the model is asked for |
+|---|---:|---|
+| `A_full_plan` | 20,940 | The complete 20-section plan |
+| `B_method_selection` | 11,896 | Pick the primary method and rule out the leading alternatives |
+| `C_completion` | 8,414 | Name what a partial plan is missing and supply it |
+| `D_critique` | 6,816 | Find the flaw in a proposed plan and fix it |
+| `H_alternatives` | 6,097 | A preferred analysis, one alternative, and when to switch |
+
+`A_full_plan` targets run to twenty numbered sections: research question, hypothesis, estimand, null/alternative, primary endpoint, analysis units, primary method, why that method, assumptions, diagnostics, effect size, uncertainty, alpha, sample size, missing data, multiplicity, subgroups, sensitivity, decision rule, limitations. Median target length is 2,391 characters; the 95th percentile is 6,705.
+
+The 6,816 `D_critique` prompts each carry exactly one deliberately planted error, drawn from nine failure modes that recur in real analysis plans:
+
+| Planted flaw | n | | Planted flaw | n |
+|---|---:|---|---|---:|
+| `ignores_clustering` | 1,243 | | `ignores_nesting` | 747 |
+| `ignores_pairing` | 1,029 | | `no_multiplicity_control` | 567 |
+| `ignores_censoring` | 911 | | `count_as_linear` | 367 |
+| `ordinal_as_continuous_or_dichotomized` | 879 | | `change_score_no_adjustment` | 227 |
+| `binary_as_continuous` | 846 | | | |
+
+## How the targets were produced
+
+Every assistant response is **synthesized from extracted structured facts**, not copied or paraphrased from source text. The pipeline pulls fields out of the registry record — design, randomization unit, arm count, outcome scale, timing, reported design parameters — and a rule-based reasoning engine keys the method off the combination of *outcome distribution × dependence structure × unit of analysis × estimand × comparison type*. A cluster-randomized trial with a continuous endpoint gets a linear mixed model with a random cluster intercept, plus a sentence explaining that clustering — not the raw participant count — drives precision. A single-arm study with a censored endpoint gets Kaplan-Meier description against a prespecified performance criterion, not a two-sample test.
+
+Two consequences follow, and both matter:
+
+- **No verbatim source text appears in any target.** Only extracted structured fields are reused, which is what makes the AEA and OSF portions distributable at all.
+- **These are not real statistical analysis plans.** No trial statistician wrote them. They are internally consistent, design-appropriate plans generated from a rule set. A model trained on them learns correct structure and method selection, not the judgement of an experienced SAP author.
+
+Curation ran as hard eligibility (outcome determinable, minimum length, bucket-adaptive quality floor) → `keep_score` ranking → budgeted allocation with source quotas, a per-family cap and a 20% ceiling on any single design. That rebalancing is why non-clinical studies make up 33.8% of these files against 21.4% of the 96k pre-curation pool, and why single-arm designs sit at 13.2% instead of 17.3%. Without it, two-arm clinical RCTs with continuous endpoints would swamp everything else.
+
+## Evaluation slices
+
+`challenge_sets` marks examples belonging to a hard slice, so you can score a model on the structures it is most likely to get wrong rather than on the corpus average. Slices overlap; membership is pipe-separated.
+
+| Slice | n | Slice | n |
+|---|---:|---|---:|
+| `clustered_hierarchical` | 8,563 | `preregistered_nonclinical` | 4,519 |
+| `flawed_plan_correction` | 6,816 | `unseen_domain_nonclinical` | 4,519 |
+| `survival_censoring` | 6,236 | `longitudinal_repeated` | 3,156 |
+| `observational_causal` | 5,539 | `count_outcomes` | 2,232 |
+| `ordinal_outcomes` | 5,002 | `multiple_testing` | 567 |
+| | | `industrial_ab_testing` | 219 |
+
+## Sources and licensing
+
+| Registry | Examples | Share | Licence posture |
+|---|---:|---:|---|
+| [ClinicalTrials.gov](https://clinicaltrials.gov/) | 35,872 | 66.2% | US government public domain |
+| [AEA RCT Registry](https://www.socialscienceregistry.org/) | 13,772 | 25.4% | Unclear — © MIT/AEA, reuse terms not explicitly open |
+| [OSF Registries](https://osf.io/registries) | 4,519 | 8.3% | Unclear — content licence not set or not verified |
+
+The AEA and OSF portions are what make the corpus non-clinical: economics field and lab experiments, psychology and ecology preregistrations, online A/B tests. They are also the portions with unresolved licence terms, tagged `human_review_status=license_review` on every row. Filter on `license` or `human_review_status` to keep only the public-domain ClinicalTrials.gov subset if your use case needs clean terms.
+
+Treat the corpus as **research use**. No verbatim source text is reproduced, but the licence posture of two of the three sources is unresolved rather than permissive.
+
+## Example
+
+A cluster-randomized ICU study, `ctgov_NCT05144607::A_full_plan`, abridged:
+
+**user**
+```text
+STUDY DESCRIPTION
+Objective: Patient-ventilator asynchronies can occur as a result of a mismatch between
+neural (patient) and ventilator inspiratory and expiratory phases. [...] The impact of
+the display of Pmus together with the other ventilator waveforms on the ability of
+health-care professionals to identify asynchronies has not been tested so far.
+Domain: clinical / Mechanical Ventilation Complication
+Design: cluster_randomized (randomized, parallel, masking: none)
+Population / eligibility: Inclusion Criteria: * Healthcare professionals (physicians and
+respiratory therapists) who work in intensive care units [...]
+Intervention/exposure: Muscle Pressure curve (Pmus)
+Comparator: placebo/standard care
+Primary outcome: Ability of ICU health care professionals to detect patient-ventilator
+asynchrony
+Primary timing: Immediately after the completion of the test sessions
+Number of groups/arms: 2
+Planned enrollment: 105 (actual)
+Reported design parameters: power 0.9, alpha 0.05
+
+Task: Produce a rigorous, prespecified hypothesis-testing and statistical-analysis plan
+for this study.
+```
+
+**assistant** (sections 6–8 of 20)
+```text
+## 6. Study and analysis units
+- Randomization/assignment unit: cluster
+- Observation unit: participant
+- Analysis unit: cluster
+- Dependence structure: Participants within a cluster are correlated (intracluster
+  correlation), so a random cluster intercept is included; ignoring it would understate
+  standard errors and inflate the type I error.
+
+## 7. Primary statistical method
+Linear mixed model with a random cluster effect. Model: Linear mixed model: outcome ~
+treatment (+ baseline, stratification factors) with a random intercept for cluster.
+Primary test: Fixed-effect test of the treatment term.
+
+## 8. Why the method is appropriate
+Because randomization is at the cluster level while the outcome is measured on
+individuals, observations are not independent within clusters. A linear mixed model with
+a random cluster intercept (or GEE with robust standard errors) propagates the
+intracluster correlation into the inference so the effective sample size, not the raw
+participant count, drives precision.
+```
+
+Section 14 shows the other half of the design. The record reports 90% power at α = 0.05 with n = 105 but no ICC, so the target does not invent one — it states that the power calculation must match the mixed-model analysis and that a design/analysis mismatch invalidates the nominal power. Unreported inputs are named as prespecification decisions throughout rather than filled in with plausible numbers.
+
+## What this corpus is not
+
+- **Not authoritative SAPs.** Synthesized targets, described above.
+- **Not a benchmark with ground-truth labels.** The `test` split is held-out training-distribution data. Use it to compare models against each other, as the model card does; a target here is a reasonable plan, not the only correct one.
+- **Not clinically validated.** Nothing here has been reviewed by a trial statistician or a regulator.
+- **Not multilingual.** English only.
+
+## Citation
+
+```bibtex
+@misc{khajegili2026hypoagentsft,
+  title        = {HypoAgent-SFT: A Corpus of Study Descriptions Paired with
+                  Prespecified Statistical Analysis Plans},
+  author       = {Khajegili Mirabadi, Ali},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/datasets/HypoAgent/HypoAgent-SFT}}
+}
+```
+
+Please also cite the underlying registries when you use the corresponding subsets. Questions: the repository discussion page, or ali.mirabadi@ubc.ca