Nishy11/nishy-al-biology-adaptive-dataset

language:

• —en task_categories:
• —question-answering
• —text-generation tags:
• —biology
• —gce-al
• —sri-lanka
• —adaptive-learning
• —mcq
• —qwen pretty_name: Nishy A/L Biology Adaptive Dataset ---

Nishy A/L Biology Adaptive Dataset

This repository contains Biology MCQ datasets prepared for the Nishy adaptive tutoring and assessment system for Sri Lankan G.C.E. A/L Biology.

Repository structure

Master audited dataset

• —biology_master_1500_final_audited.json
• —Final audited master collection containing 1500 MCQs.

V4 paper-level split

• —v4_train_base_1182.json
• —v4_validation_base_90.json
• —v4_test_final_76.json

These files represent the cleaned paper-level training, validation and held-out test splits.

V5 adaptive examples

• —v5_train_examples_5910.json
• —v5_validation_examples_450.json

These contain instruction-style examples prepared for adaptive tutoring and MCQ generation workflows.

Source datasets

The source_datasets/ directory contains the individual preprocessed source collections used during dataset construction.

It includes:

• —Sri Lankan G.C.E. A/L Biology past-paper collections
• —School and provincial Biology MCQ papers
• —500-MCQ tutorial batches
• —Supportive assessment collections
• —Synthetic hard Biology MCQ sets

These source files are retained separately to support dataset auditing, provenance tracking and reproducibility.

Adaptive tutoring design

The Nishy tutoring workflow is designed as:

1. 1.First wrong attempt -> HARD Socratic hint
2. 2.Second wrong attempt -> MEDIUM hint
3. 3.Third wrong attempt -> EASY hint
4. 4.Fourth wrong attempt -> Correct answer with explanation

Hints are intended to guide biological reasoning without directly revealing the answer before the final attempt.

Dataset preparation

Processing included:

• —option-format normalization
• —answer-key auditing where source information was available
• —duplicate analysis
• —eligibility filtering
• —paper-level train / validation / test separation
• —adaptive hint example generation
• —direct answer-leak auditing

Model development

Base model:

Qwen/Qwen2.5-7B-Instruct

Fine-tuned adapter:

Nishy11/nishy-qwen2.5-7b-v4

The adaptive application also uses inference-time orchestration and validation logic in addition to the fine-tuned adapter.

Intended use

This dataset is intended for:

• —adaptive Biology tutoring
• —MCQ generation
• —MCQ answer reasoning
• —Socratic hint generation
• —explanation generation
• —educational NLP research
• —instruction tuning of language models

Limitations

• —Source collections differ in wording, difficulty and provenance.
• —Generated hints require pedagogical and answer-leak validation.
• —The dataset should not be treated as an official examination resource.
• —Redistribution rights for source-derived examination and tutorial material should be reviewed before making the repository public.

Important licensing note

No blanket license is currently asserted for all source-derived files in this repository.

Some source material may originate from examination papers, schools, tutorial collections or other third-party educational resources. Copyright and redistribution permissions should be verified before changing this dataset repository from private to public.

Disclaimer

This is an independently prepared educational/research dataset and is not an official publication of the Sri Lankan Department of Examinations, NIE, or any listed school or educational institution.