Hugging Face's logo

Datasets:
ssfc
/
pcf-method-delta-sft-data

Tasks:
Text Generation
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
Tags:
algorithms
supervised-fine-tuning
method-delta
idea-generation
research-assistant
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
pcf-method-delta-sft-data / README.md
ssfc's picture
ssfc
Upload PCF method-delta SFT datasets
08ccfa9 verified
preview code
|
Raw History Blame Contribute Delete
2.85 kB
metadata 
pretty_name: PCF Method Delta SFT Data
language:
  - en
task_categories:
  - text-generation
tags:
  - algorithms
  - supervised-fine-tuning
  - method-delta
  - idea-generation
  - research-assistant
license: other

PCF Method Delta SFT Data

This dataset contains supervised fine-tuning examples for Past Creates Future (PCF), a research workflow for generating algorithmic ideas by learning method-level deltas between related algorithms.

Each example is a JSON object with:

  • prompt: an instruction plus an existing algorithm description and a method delta.
  • response: the target improved algorithm description in JSON form.

Files

File Rows Purpose
pcf_detail_to_detail_30k_conf0p6.jsonl 5,067 Main dataset used for detail-to-detail SFT. The prompt and response use richer method-detail fields.
pcf_compact_30k_conf0p6.jsonl 5,067 Compact baseline dataset. The target focuses on novelty_claim, core_idea, and algorithm_description.

Intended Use

The dataset is intended for research on:

  • algorithm idea generation;
  • method-delta transfer;
  • supervised fine-tuning of research assistant models;
  • comparing compact algorithm generation against richer detail-to-detail generation.

The dataset is not intended to reproduce source papers verbatim. It contains structured training examples derived from method summaries and method-delta summaries.

Provenance

The examples were generated from a local corpus of algorithm papers using a multi-stage PCF preprocessing pipeline:

  1. Extract method summaries from paper text.
  2. Extract or summarize method-level deltas between related algorithm pairs.
  3. Combine existing-method detail, delta detail, and target-method detail into SFT prompt/response examples.
  4. Filter examples using confidence thresholds and manually maintained exclusion lists for non-method or problematic inputs.

The main detail-to-detail dataset corresponds to the 30k extraction setting and confidence threshold 0.6.

Limitations

  • The dataset is derived from automatically extracted summaries, so individual examples may contain extraction or summarization errors.
  • Confidence filtering reduces but does not eliminate noisy pairs.
  • The dataset is designed for generating candidate research ideas; empirical algorithm performance must still be validated by downstream experiments.
  • Licensing of the original papers varies. This dataset is released as structured derived summaries for research use and does not include paper PDFs, full paper text, GROBID XML, or other raw paper artifacts.

Related Models

Models trained from this data include:

  • ssfc/pcf-qwen3-14b-detail-to-detail-lora
  • ssfc/pcf-qwen3.5-9b-detail-to-detail-lora
  • ssfc/pcf-ministral3-base-detail-to-detail-lora
  • ssfc/pcf-qwen3-14b-compact-30k-conf0p6-lora