Qwen3-8B-SPoT

Model Description

Qwen3-8B-SPoT is a reasoning-enhanced large language model post-trained from the Qwen/Qwen3-8B base model. It is trained using the Surgical Post-Training (SPoT) paradigm, which significantly improves the model's reasoning capabilities while alleviating the catastrophic forgetting typically associated with Supervised Fine-Tuning (SFT).

This model was introduced in the paperSurgical Post-Training: Cutting Errors, Keeping Knowledge (Lin & Han, 2026).

• Code Repository: Visual-AI/SPoT

Training Details & Performance

• Efficiency: The model was trained using merely 4k rectified math data pairs. It avoids standard multi-phase pipelines (SFT → GRPO → DPO).
• Reasoning Improvement: SPoT improves the base Qwen3-8B's accuracy by 6.2% on average across in-domain and Out-of-Domain (OOD) complex math and reasoning tasks.
• Knowledge Retention: The model robustly mitigates catastrophic forgetting, yielding stability on general capability benchmarks like IFEval.

Paper Results

Results reported in the SPoT paper for the Qwen3-8B setting:

Method	In-domain Avg	OOD Avg	IFEval	Overall Avg
Qwen3-8B (base)	46.8	29.9	83.0	47.1
+ SFT	41.0 (-5.8)	25.5 (-4.4)	79.6 (-3.4)	41.8 (-5.3)
+ RFT	47.3 (+0.5)	26.1 (-3.8)	81.5 (-1.5)	46.4 (-0.7)
+ SFT+	50.5 (+3.7)	30.7 (+0.8)	80.0 (-3.0)	49.4 (+2.3)
+ SPoT (this model)	52.1 (+5.3)	41.4 (+11.5)	84.8 (+1.8)	53.3 (+6.2)

In-domain Avg covers AIME24/25, AMC23, MATH-500, Minerva, and OlympiadBench. OOD Avg covers GPQA-Diamond and Connect4. IFEval measures instruction following.

Usage

You can load and generate text with this model using the standard Hugging Face transformers pipeline:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "linius/Qwen3-8B-SPoT"

# Load Tokenizer and Model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Prepare prompt
prompt = "Solve the following math problem step-by-step: ..."
messages =[
    {"role": "system", "content": "You are a helpful and precise reasoning assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048
)
generated_ids =[
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Citation

If you find our model, data, or the SPoT methodology useful in your research, please consider citing our paper:

BibTeX:

@article{lin2026surgical,
      title={Surgical Post-Training: Cutting Errors, Keeping Knowledge}, 
      author={Wenye Lin and Kai Han},
      year={2026},
      journal={arXiv preprint arXiv:2603.01683}
}