language:
- en
pretty_name: CODA-19 - Scientific Abstract Segmentation and Classification
tags:
- coda-19
- scientific-abstracts
- segmentation
- classification
configs:
- config_name: default
data_files:
- split: train
path: coda19_train_6000_final_sft.jsonl.gz
- split: test
path: coda19_test_400_final_sft.jsonl
- config_name: dev181
data_files:
- split: validation
path: coda19_dev_200_max25.jsonl
CODA-19 – Scientific Abstract Segmentation and Classification
This dataset contains the training, test and development data used for a Qwen3.5-4B QLoRA fine-tuning project on scientific abstract segmentation and classification.
The model segments complete scientific abstracts into consecutive text spans, assigns one of five research aspects (background, purpose, method, finding, other), and generates a reference to the corresponding codebook rule or definition.
Dataset Splits
| File | Abstracts | Split | Configuration |
|---|---|---|---|
coda19_train_6000_final_sft.jsonl.gz |
6,000 | train | default |
coda19_test_400_final_sft.jsonl |
400 | test | default |
coda19_dev_200_max25.jsonl |
181 | validation | dev181 |
The development data originates from a selection of 200 abstracts, of which 181 are included in the provided file and evaluation.
Train and test contain structured SFT examples with paper_id and messages. The development file contains abstract texts, gold segments and additional metadata and therefore uses a separate dataset configuration.
The test set was also used to monitor evaluation loss and select the training checkpoint. The development set was not used for checkpoint selection.
Data Provenance
The original scientific abstracts, segment boundaries and research-aspect labels originate from CODA-19 (Huang et al., 2020), based on CORD-19.
The original human annotations were retained. Codebook references were added with LLM assistance to provide additional structured training targets.
Project Repository
https://github.com/mc990-ai/CODA19-Qwen3.5-4B
Fine-Tuned Model
https://huggingface.co/mc9901708/CODA19-Qwen3.5-4B-LoRA
Rights and Licensing
The original abstracts originate from scientific publications with potentially different licensing conditions. No blanket license for the underlying abstracts is granted by this dataset repository. Rights remain with the respective rights holders.
See the GitHub repository for detailed documentation, training methodology and evaluation results.