Datasets:
| language: | |
| - en | |
| license: mit | |
| task_categories: | |
| - question-answering | |
| - text2text-generation | |
| tags: | |
| - biomedical | |
| - pubmedqa | |
| - medquad | |
| - evidence-based-medicine | |
| pretty_name: BioEvidence-SFT | |
| size_categories: | |
| - 1K<n<10K | |
| # Dataset Card: BioEvidence-SFT & BioEvidence-Eval | |
| ## Dataset Summary | |
| The **BioEvidence-SFT** and **BioEvidence-Eval** datasets are curated instruction-tuning and benchmarking corpora designed for evidence-grounded biomedical NLP tasks. | |
| - **GitHub Repository (Code & Pipeline):** [https://github.com/PRATHMESHTALHANDE/BioEvidence-LLM-4bit-QLoRA](https://github.com/PRATHMESHTALHANDE/BioEvidence-LLM-4bit-QLoRA) | |
| - **Model Adapter:** [https://huggingface.co/Bhupati1998/BioEvidence-LLM-1.5B](https://huggingface.co/Bhupati1998/BioEvidence-LLM-1.5B) | |
| ## Subsets & Splits | |
| - **BioEvidence-SFT-v0.1:** 100 balanced pilot samples across all 7 taxonomy tasks (100% human-rubric validated). | |
| - **BioEvidence-SFT-full:** 880 multi-source training samples formatted in ChatML format. | |
| - **BioEvidence-Eval-v0.1:** 156 held-out evaluation samples with zero article/PMID overlap with training data. | |
| ## Provenance & Sources | |
| 1. **PubMedQA (MIT License):** Expert-annotated research questions with `yes`/`no`/`maybe` decisions. | |
| 2. **MedQuAD (NIH Public Domain):** 12 NIH institute consumer and clinician health resources. | |
| 3. **PubMed (NCBI E-utilities):** Controlled queries for randomized clinical trials. | |
| 4. **PMC Open Access (CC-BY):** Full-text open-access articles via the BioC API. | |
| ## Leakage Prevention | |
| All splits are grouped strictly by PubMed ID (`PMID`) or PubMed Central ID (`PMCID`). No articles or abstracts exist in both training and evaluation subsets. | |