Miner-4B
This repository hosts the Hugging Face Transformers checkpoint for MINER: Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Models.
Model Description
Miner-4B is a reasoning model trained with MINER, a reinforcement learning method designed to improve data efficiency for large reasoning models. MINER targets the inefficiency of critic-free RL methods on positive homogeneous prompts, where all sampled rollouts are correct and standard relative-advantage training provides little or no learning signal. Instead, MINER leverages the policy’s intrinsic uncertainty as a self-supervised reward signal, without requiring auxiliary reward models or additional inference-time overhead. :contentReference[oaicite:1]{index=1}
The MINER framework introduces two central ideas:
- Token-level focal credit assignment, which amplifies learning on uncertain and critical tokens while suppressing overconfident ones.
- Adaptive advantage calibration, which integrates intrinsic and verifiable rewards in a stable way. :contentReference[oaicite:2]{index=2}
According to the paper, MINER is evaluated on six reasoning benchmarks using Qwen3-4B-Base and Qwen3-8B-Base, and reports stronger sample efficiency and accuracy than several baseline methods including GRPO variants. :contentReference[oaicite:3]{index=3}
Intended Use
This model is intended for research and experimental use in:
- reasoning and problem solving
- reinforcement learning for language models
- mathematical and verifiable reasoning tasks
- post-training and evaluation of large reasoning models
Potential use cases include:
- academic research on RL for reasoning models
- evaluation on reasoning benchmarks
- ablation and reproduction studies based on the MINER framework
- further finetuning or post-training from this checkpoint
How to Use
Transformers
1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3model_name = "pixas/Miner-4B"
4
5tokenizer = AutoTokenizer.from_pretrained(model_name)
6model = AutoModelForCausalLM.from_pretrained(
7 model_name,
8 torch_dtype="auto",
9 device_map="auto"
10)
11
12prompt = [{"role": "user", "content": "What is 2+3?"}]
13inputs = tokenizer(tokenizer.apply_chat_template(prompt, add_generation_prompt=True, tokenize=False), return_tensors='pt').to(model.device)
14
15outputs = model.generate(
16 **inputs,
17 max_new_tokens=8192,
18 do_sample=True
19)
20
21print(tokenizer.decode(outputs[0], skip_special_tokens=True))
vLLM
1from vllm import LLM, SamplingParams
2
3llm = LLM(model="pixas/Miner-4B")
4sampling_params = SamplingParams(
5 temperature=0.6,
6 max_tokens=8192
7)
8prompt = [{"role": "user", "content": "What is 2+3?"}]
9inputs = tokenizer.apply_chat_template(prompt, add_generation_prompt=True, tokenize=False)
10outputs = llm.generate(
11 inputs,
12 sampling_params
13)
14
15print(outputs[0].outputs[0].text)
Limitations
This model is a research checkpoint and may have several limitations:
- It may produce incorrect, incomplete, or overconfident reasoning outputs.
- Performance may depend heavily on prompt format and decoding setup.
- Results reported in the paper may not transfer exactly to this released checkpoint unless the same base model, data mixture, and evaluation pipeline are used.
- The model is not intended as a substitute for expert judgment in high-stakes domains.
Bias, Risks, and Safety
Like other large language models, this model may reflect biases present in its training data and may generate harmful, misleading, or factually incorrect outputs. Additional care is required before deployment in user-facing or safety-critical applications.
Citation
If you use this model, please cite:
1@article{jiang2026miner,
2 title={Miner: Mining Intrinsic Mastery for Data-Efficient RL in Large Reasoning Models},
3 author={Jiang, Shuyang and Wang, Yuhao and Zhang, Ya and Wang, Yanfeng and Wang, Yu},
4 journal={arXiv preprint arXiv:2601.04731},
5 year={2026}
6}
Acknowledgements
This model card is based on the official MINER paper and code repository:
