---
license: other
language:
- en
pretty_name: SLM-in-SciPaper
task_categories:
- token-classification
- text-classification
- summarization
- question-answering
configs:
- config_name: training_manifest
  data_files:
  - split: train
    path: data/training_datasets_manifest.csv
- config_name: keyword_kp20k
  data_files:
  - split: train
    path: data/keyword_keyphrase/kp20k/train.jsonl
  - split: validation
    path: data/keyword_keyphrase/kp20k/validation.jsonl
  - split: test
    path: data/keyword_keyphrase/kp20k/test.jsonl
- config_name: keyword_ldkp10k
  data_files:
  - split: train
    path: data/keyword_keyphrase/ldkp10k/train.jsonl
  - split: validation
    path: data/keyword_keyphrase/ldkp10k/validation.jsonl
  - split: test
    path: data/keyword_keyphrase/ldkp10k/test.jsonl
- config_name: keyword_semeval2010
  data_files:
  - split: train
    path: data/keyword_keyphrase/semeval2010/train.jsonl
  - split: test
    path: data/keyword_keyphrase/semeval2010/test.jsonl
- config_name: structure_pubmed_rct
  data_files:
  - split: train
    path: data/structure/pubmed_rct/train.jsonl
  - split: validation
    path: data/structure/pubmed_rct/validation.jsonl
  - split: test
    path: data/structure/pubmed_rct/test.jsonl
- config_name: structure_qasper
  data_files:
  - split: train
    path: data/structure/qasper/train.jsonl
  - split: validation
    path: data/structure/qasper/validation.jsonl
  - split: test
    path: data/structure/qasper/test.jsonl
- config_name: structure_aclsum
  data_files:
  - split: train
    path: data/structure/aclsum/train.jsonl
  - split: validation
    path: data/structure/aclsum/validation.jsonl
  - split: test
    path: data/structure/aclsum/test.jsonl
- config_name: paper_corpus_manifest
  data_files:
  - split: train
    path: data/paper_corpus/manifest.csv
---

# SLM-in-SciPaper

This repository stores the data and model assets used by the SLM-in-SciPaper project.

## Contents

- `data/keyword_keyphrase`: processed Stage 1 keyphrase extraction data.
- `data/structure`: processed Stage 2 structural evidence modeling data.
- `data/paper_corpus/full_library_txt`: 178 plain-text scientific papers used as the local demonstration corpus.
- `data/paper_corpus/manifest.csv`: metadata and file paths for the 178-paper text corpus.
- `data/training_datasets_manifest.csv`: a compact index of the processed training/evaluation files.
- `resources/lexicon`: supplementary section BoW, TF-IDF, frequency, and evidence cue resources used by the inference pipeline.
- `models`: model assets/checkpoints previously uploaded for this project.

## Training Data Actually Used

The final local pipeline used six processed training datasets:

| Stage | Dataset | Role in the project |
|---|---|---|
| Stage 1 | KP20k | keyword extractor warm-up |
| Stage 1 | LDKP10K | long-document keyword fine-tuning |
| Stage 1 | SemEval2010 | final keyword fine-tuning/evaluation |
| Stage 2 | PubMed RCT | hard role supervision |
| Stage 2 | QASPER | evidence supervision and partial role constraint |
| Stage 2 | ACLSum | importance supervision and facet role supervision |

SciTLDR weak data existed in the local workspace but was not used by the final reported training pipeline, so it is not included as a training dataset here.

## Data Format

The processed training data is stored as JSONL files. Each line is one document-level record in the schema consumed by the project code. Summary files are kept beside each dataset directory when available.

The paper corpus is stored as individual `.txt` files. Use `data/paper_corpus/manifest.csv` to map each paper id to its text file.

## Notes

This repository contains processed versions of public research datasets and a local paper-text corpus for course/research use. Original upstream dataset licenses and terms should be checked before redistribution or commercial use.