usama10/grpo-tax-eval-data

GRPO Tax: Evaluation Data

Complete dense-checkpoint evaluation data from the paper:

The GRPO Tax is Smaller Than You Think: A Longitudinal Study of Capability Preservation During Reasoning Training Muhammad Usama. Transactions on Machine Learning Research, 2026. OpenReview | Code

Contents

432 JSON evaluation files holding 5,616 individual benchmark scores, from dense-checkpoint evaluation of 5 models during GRPO training and 2 models during DPO training.

qwen-1.5b/          # 76 files (base + 75 checkpoints)
qwen-3b/            # 76 files
phi-3.8b/           # 76 files
gemma-2b/           # 76 files
llama-3b/           # 76 files
qwen-1.5b-dpo/      # 26 files (base + 25 checkpoints)
qwen-3b-dpo/        # 26 files

Download

This is a model-type repository, so no --repo-type flag is needed:

pip install huggingface_hub
hf download usama10/grpo-tax-eval-data --local-dir results/

To regenerate every figure in the paper from this data:

git clone https://github.com/Usama1002/GRPO-Capability-Tax.git
cd GRPO-Capability-Tax
hf download usama10/grpo-tax-eval-data --local-dir results/
python analysis/generate_all_figures.py

File Format

Each file holds scores on 13 benchmarks for a single checkpoint:

{
  "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
  "checkpoint": "results/qwen-1.5b/training/checkpoint-3736",
  "results": {
    "math_reasoning": {"benchmark": "math_reasoning", "metric": "accuracy", "score": 0.355, "num_examples": 200}
  }
}

Benchmarks

Benchmark	Metric	Examples
GSM8K (target)	Accuracy	200
MMLU	Accuracy	200
HellaSwag	Accuracy	200
ARC-Challenge	Accuracy	200
TruthfulQA	Accuracy	200
Winogrande	Accuracy	200
IFEval	Constraint satisfaction	30
XSum	ROUGE-L	66
WMT (en->de)	BLEU	97
Coding	Syntax + structure	25
Safety	Refusal rate	30
Creative Writing	Vocabulary diversity	25
Conversation	Helpfulness	25

The five heuristic benchmarks (IFEval, Coding, Safety, Creative Writing, Conversation) use rule-based scoring and are approximate; the paper reports expanded-sample-size audits for them.

Related Resources

Resource	Link
Paper	The GRPO Tax is Smaller Than You Think: A Longitudinal Study of Capability Preservation During Reasoning Training (TMLR, 2026)
Source code	github.com/Usama1002/GRPO-Capability-Tax
Evaluation data	usama10/grpo-tax-eval-data
GRPO adapters	qwen-1.5b, qwen-3b, phi-3.8b, gemma-2b, llama-3b
DPO adapters	qwen-1.5b-dpo, qwen-3b-dpo

Citation

@article{usama2026grpotax,
  title   = {The {GRPO} Tax is Smaller Than You Think: A Longitudinal Study of Capability Preservation During Reasoning Training},
  author  = {Muhammad Usama},
  journal = {Transactions on Machine Learning Research},
  issn    = {2835-8856},
  year    = {2026},
  url     = {https://openreview.net/forum?id=e0UVcimXdK}
}