Xiaofeng77/diverse_sokoban

Debunk the Myth of SFT Generalization Dataset

This dataset is associated with the paper Debunk the Myth of SFT Generalization, which re-evaluates the generalization capabilities of supervised fine-tuning (SFT) compared to reinforcement learning (RL) on decision-making benchmarks. The research demonstrates that with proper data curation, such as prompt diversity and Chain-of-Thought (CoT) supervision, SFT can achieve strong generalization, matching or even surpassing RL baselines.

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

Abstract

A prevailing view holds that supervised fine-tuning (SFT) memorizes training data and fails to generalize, whereas reinforcement learning (RL) attains broader robustness. We revisit this claim through a systematic evaluation on two decision-making benchmarks, Sokoban and General Points, and arrive at a different conclusion. We show that much of SFT's perceived failure stems from frozen-prompt artifacts: when trained on fixed instruction templates, SFT models cling to training semantics rather than adapting to new ones. Introducing prompt diversity during training breaks this shortcut and yields strong generalization to unseen instruction variants without harming in-distribution performance. Beyond instruction shifts, we ask whether SFT can generalize to strictly harder tasks. Here, chain-of-thought (CoT) supervision provides an algorithmic scaffold that markedly improves transfer to more difficult regimes, such as larger Sokoban grids with additional boxes and arithmetic with out-of-distribution values or five-card compositions that increase combinatorial complexity. Finally, combining prompt diversity with CoT achieves the best of both worlds: robust generalization across both instruction-variant and difficulty-variant settings, matching or surpassing RL baselines on our benchmarks while retaining SFT's simplicity and stability. These findings challenge the narrative that SFT is inherently inferior to RL and support a data-centric perspective: with appropriately curated demonstrations, vanilla SFT can generalize as strongly as RL. Code reproducing the results in the paper can be found at: this https URL .

Sample Usage

The associated GitHub repository provides scripts to train models using the datasets. Ensure you have the necessary environment set up as described in the repository's Installation section before running these commands.

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