---
language:
- en
license: other
task_categories:
- text-classification
- text-generation
pretty_name: Cochrane Screening SFT
tags:
- systematic-review
- cochrane
- title-abstract-screening
- medical
- peft
- sft
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: val.jsonl
  - split: test
    path: test.jsonl
- config_name: heldout_reviews1
  data_files:
  - split: test
    path: heldout_reviews1.jsonl
- config_name: heldout_reviews2
  data_files:
  - split: test
    path: heldout_reviews2.jsonl
- config_name: heldout_reviews3
  data_files:
  - split: test
    path: heldout_reviews3.jsonl
dataset_info:
- config_name: default
  features:
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: row_id
    dtype: int64
  - name: label
    dtype: string
  splits:
  - name: train
    num_examples: 416799
  - name: validation
    num_examples: 46311
  - name: test
    num_examples: 26673
- config_name: heldout_reviews1
  features:
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: row_id
    dtype: int64
  - name: label
    dtype: string
  splits:
  - name: test
    num_examples: 26858
- config_name: heldout_reviews2
  features:
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: row_id
    dtype: int64
  - name: label
    dtype: string
  splits:
  - name: test
    num_examples: 4391
- config_name: heldout_reviews3
  features:
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: row_id
    dtype: int64
  - name: label
    dtype: string
  splits:
  - name: test
    num_examples: 19007
---

# Cochrane Screening SFT

Supervised fine-tuning (SFT) chat dataset for **Cochrane-style title and abstract screening**.

Each example is a chat conversation that asks a model to predict a screening decision
(`include` / `exclude` / `uncertain`) and a short justification (`reason`).

Code: [ljwa2323/cochrane-screening-slm](https://github.com/ljwa2323/cochrane-screening-slm)

## Dataset summary

| Split / config | Records | Role |
| --- | ---: | --- |
| `train` | 416,799 | LoRA SFT training |
| `validation` | 46,311 | Training-time validation (10% stratified holdout from development data) |
| `test` | 26,673 | Internal held-out test split |
| `heldout_reviews1` | 26,858 | External reviews (random Cochrane set) |
| `heldout_reviews2` | 4,391 | External reviews (HIV-focused set) |
| `heldout_reviews3` | 19,007 | External reviews (heart/CVD-focused set) |

Label mapping used when building the dataset:

- `0.0` -> `exclude`
- `0.5` -> `uncertain`
- `1.0` -> `include`

Approximate label counts on the development-derived set (train+val source):

- exclude: 221,129
- uncertain: 160,861
- include: 81,120

## Data fields

Each `*.jsonl` line contains:

| Field | Type | Description |
| --- | --- | --- |
| `messages` | list | Chat turns: `system`, `user`, `assistant` |
| `row_id` | int | Source row id |
| `label` | string | Gold label: `include` / `exclude` / `uncertain` |

The assistant target is a JSON object:

```json
{"label": "include|exclude|uncertain", "reason": "<brief explanation>"}
```

## How to load

```python
from datasets import load_dataset

# Internal splits
ds = load_dataset("deepcoder2024/cochrane-screening-sft")
print(ds)

# External held-out reviews
hr1 = load_dataset("deepcoder2024/cochrane-screening-sft", "heldout_reviews1")
hr2 = load_dataset("deepcoder2024/cochrane-screening-sft", "heldout_reviews2")
hr3 = load_dataset("deepcoder2024/cochrane-screening-sft", "heldout_reviews3")
```

> Tip: ignore `*_manifest.json` files when loading. They are metadata only and should not be parsed as chat examples.

## Intended use

- Fine-tune small language models (e.g., Qwen3 LoRA) for title/abstract screening
- Evaluate screening label + reason generation on internal and external review sets

## Out-of-scope use

- Not a substitute for expert systematic-review judgment
- Not intended for clinical decision-making about individual patients
- Labels and reasons are for research / screening-assistance experiments only

## Related models

- [`deepcoder2024/Qwen3-1.7B-LoRA-Cochrane-Screening`](https://huggingface.co/deepcoder2024/Qwen3-1.7B-LoRA-Cochrane-Screening)
- [`deepcoder2024/Qwen3-4B-LoRA-Cochrane-Screening`](https://huggingface.co/deepcoder2024/Qwen3-4B-LoRA-Cochrane-Screening)
- [`deepcoder2024/Qwen3-8B-LoRA-Cochrane-Screening`](https://huggingface.co/deepcoder2024/Qwen3-8B-LoRA-Cochrane-Screening)

## Citation

If you use this dataset, please cite the associated project repository:

```bibtex
@misc{cochrane_screening_sft,
  title        = {Cochrane Screening SFT Dataset},
  author       = {deepcoder2024},
  year         = {2026},
  howpublished = {\\url{https://huggingface.co/datasets/deepcoder2024/cochrane-screening-sft}}
}
```