README.md

metadata

task_categories:
  - TEXT_GENERATION
language:
  - en
tags:
  - supervised-fine-tuning
  - reinforcement-learning
  - sokoban
  - general-points
  - chain-of-thought
  - instruction-following
  - reasoning
  - decision-making
  - llm
dataset_info:
  features:
    - name: data_source
      dtype: string
    - name: prompt
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
    - name: ability
      dtype: string
    - name: reward_model
      struct:
        - name: ground_truth
          sequence: int64
        - name: style
          dtype: string
    - name: extra_info
      struct:
        - name: index
          dtype: int64
        - name: split
          dtype: string
    - name: id
      dtype: string
  splits:
    - name: train
      num_bytes: 3590654
      num_examples: 3982
    - name: test
      num_bytes: 1645357
      num_examples: 1602
    - name: test_env
      num_bytes: 115290
      num_examples: 100
  download_size: 877296
  dataset_size: 5351301
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: test_env
        path: data/test_env-*

Debunk the Myth of SFT Generalization Dataset

This dataset is associated with the paper Debunk the Myth of SFT Generalization.

The research presented in the paper challenges the common belief that supervised fine-tuning (SFT) primarily memorizes training data and lacks generalization, especially when compared to reinforcement learning (RL). Through systematic evaluation on decision-making benchmarks like Sokoban and General Points, the authors demonstrate that incorporating prompt diversity and Chain-of-Thought (CoT) supervision during SFT training can lead to strong generalization across unseen instruction variants and strictly harder tasks, often matching or surpassing RL baselines.

Code: https://github.com/XiaofengLin7/debunking-sft-generalization

Installation

This dataset is intended for use with the associated code repository. To set up the environment and dependencies:

Prerequisites

CUDA 12.2 & cuDNN 9.1.0 works, but official docs recommends CUDA >= 12.4 & cuDNN >= 9.8.0.

Setup

conda create -n debunk_sft python=3.10
conda activate debunk_sft
USE_MEGATRON=0 bash setup.sh
git submodule init
git submodule update
pip install -e thirdparty/verl --no-dependencies
pip install -e thirdparty/ragen --no-dependencies
pip install -e thirdparty/alfworld --no-dependencies
pip install -e thirdparty/trl --no-dependecies

Dataset Overview

This repository contains various datasets used in the research, categorized by task, method, diversity, and format. These datasets are part of a larger collection.

Task	Method	Diversity	Format	Link
Sokoban	RL	non-diverse	—	🤗
Sokoban	RL	diverse	—	🤗
Sokoban	SFT	non-diverse	answer-only	🤗
Sokoban	SFT	diverse	answer-only	🤗
Sokoban	SFT	non-diverse	cot	🤗
Sokoban	SFT	diverse	cot	🤗
General Points	RL	non-diverse	—	🤗
General Points	RL	diverse	—	🤗
General Points	SFT	non-diverse	answer-only	🤗
General Points	SFT	diverse	answer-only	🤗
General Points	SFT	non-diverse	cot	🤗
General Points	SFT	diverse	cot	🤗

Sample Usage

The following snippets from the GitHub repository demonstrate how to train models using this dataset with SFT (Supervised Fine-Tuning) or GRPO (Generative Reinforcement Policy Optimization).

Train your model with SFT

Specify your model and data beforehand.

For Sokoban:

bash debunk_sft/scripts/sokoban/sokoban_train_and_eval.sh

For General Points:

bash debunk_sft/scripts/gp_l/gp_l_train_and_eval.sh

Train your model with GRPO

Specify your model and data beforehand.

For Sokoban:

bash debunk_sft/scripts/sokoban/sokoban_grpo.sh

For General Points:

bash debunk_sft/scripts/gp_l/gp_l_grpo.sh

Citation

If you use this dataset in your research, please cite the associated paper:

@misc{lin2024debunkthemythofsftgeneralization,
      title={Debunk the Myth of SFT Generalization},
      author={Xiaofeng Lin and Yuandong Tian and Huazhe Xu},
      year={2024},
      eprint={2510.00237},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.00237},
}