Instructions to use CMU-AIRe/MRT-offline with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use CMU-AIRe/MRT-offline with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="CMU-AIRe/MRT-offline")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("CMU-AIRe/MRT-offline")
model = AutoModelForCausalLM.from_pretrained("CMU-AIRe/MRT-offline")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use CMU-AIRe/MRT-offline with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "CMU-AIRe/MRT-offline"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-offline",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/CMU-AIRe/MRT-offline

SGLang

How to use CMU-AIRe/MRT-offline with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "CMU-AIRe/MRT-offline" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-offline",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "CMU-AIRe/MRT-offline" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-offline",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use CMU-AIRe/MRT-offline with Docker Model Runner:
```
docker model run hf.co/CMU-AIRe/MRT-offline
```

MRT-offline (R1-Distill-Qwen-1.5B)

DeepSeek-R1-Distill-Qwen-1.5B fine-tuned with Meta Reinforcement Fine-Tuning (MRT) in the open-ended setting — the paper-faithful "offline" variant: a dense progress reward (the change in the likelihood of eventual success contributed by each reasoning episode), estimated from a forced-termination meta-prover over an off-policy prefix and applied as a single end-of-trace bonus on top of outcome-reward GRPO.

📄 Paper: Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
💻 Code / recipe: CMU-AIRe/MRT → train/rl/ (REPRODUCTION.md)
🔀 On-policy variant: CMU-AIRe/MRT-online

v0.1 reproduction. This checkpoint was produced by the open-source v0.1 training code (built on miles, since the original Open-R1 setup is no longer actively maintained), not the exact run from the paper. It reproduces the paper's relative claim — MRT's gain over the base is ~2–3× the gain from outcome-reward GRPO — at a slightly smaller absolute magnitude (see below).

Evaluation

pass@1 (mean of 64 samples/problem) at a 16K token budget, averaged over AIME 2024 / AIME 2025 / AMC 2023 / MinervaMATH / MATH500 (single grader used for training and eval):

model	AIME24	AIME25	AMC23	Minerva	MATH500	Avg	gain over base
base (R1-Distill-Qwen-1.5B)	27.34	22.86	67.89	24.94	81.71	44.95	—
GRPO (outcome-reward)	28.12	22.97	67.77	26.45	81.85	45.43	+0.48
MRT-offline (this model)	28.75	23.59	70.86	24.96	82.61	46.16	+1.20

MRT's gain over base (+1.20) is 2.5× the GRPO gain (+0.48) — within the paper's reported 2–3× range. (Paper Table 1 reports +1.1 / +2.2 for GRPO / MRT; our reproduction reaches a smaller absolute magnitude — see REPRODUCTION.md for the training-length and grader notes.)

Training

Base: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B; data: 4,000 NuminaMath problems.
GRPO + α-weighted progress bonus (α=1.0), 248 optimizer steps, 16K budget, temp 0.9.
Framework: miles (Megatron-LM + SGLang). Full recipe, hyperparameters, and assumptions: CMU-AIRe/MRT train/rl/REPRODUCTION.md.

Citation

@misc{qu2025optimizingtesttimecomputemeta,
      title={Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning},
      author={Yuxiao Qu and Matthew Y. R. Yang and Amrith Setlur and Lewis Tunstall and Edward Emanuel Beeching and Ruslan Salakhutdinov and Aviral Kumar},
      year={2025}, eprint={2503.07572}, archivePrefix={arXiv}, primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.07572},
}

Downloads last month: 1

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for CMU-AIRe/MRT-offline

Base model

deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

Finetuned

(640)

this model

Paper for CMU-AIRe/MRT-offline

Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning

Paper • 2503.07572 • Published Mar 10, 2025 • 48