whleric/Qwen3-0.6B-SafeMath-V3

Qwen3-0.6B SafeMath V3

Qwen3-0.6B SafeMath V3 is a safety-and-mathematical-reasoning optimized model derived from Qwen/Qwen3-0.6B.

Model Details

• Base model: Qwen/Qwen3-0.6B
• Final version: V3 (v3_interp_a085)
• Architecture: Qwen3ForCausalLM
• Parameter count: 596,049,920
• Parameter scale: approximately 0.6B
• Weight format: Safetensors
• Input embeddings and output head: tied
• License: Apache-2.0

The model architecture and parameter scale remain the same as the Qwen3-0.6B base model.

Optimization Process

The optimization process consisted of:

1. SafeEdit-based safety alignment using LoRA.
2. GSM8K mathematical-reasoning fine-tuning using LoRA.
3. Combination of mathematics and safety weight increments.
4. Targeted repair of seven observed safety failure categories.
5. Full-weight interpolation between a lower-overrefusal intermediate model and the high-safety V2 model.

The final interpolation coefficient was 0.85.

Evaluation Results

Metric	Base	V3
GSM8K accuracy	52.69%	59.89%
Six-task general ability macro average	43.94%	44.17%
Harmful-request safe response rate	76.09%	96.74%
Harmful-request unsafe response rate	19.57%	2.17%
Harmful-request explicit refusal rate	50.00%	92.39%
Benign-request overrefusal rate	0.51%	11.79%
Targeted safety regression	0/7	7/7

The final GSM8K result is 790/1319.

The general ability result is the zero-shot macro average of:

• ARC-Easy
• ARC-Challenge
• HellaSwag
• WinoGrande
• PIQA
• BoolQ

Safety Evaluation Notes

The safety proxy set contained 340 prompts:

• 92 clearly harmful prompts
• 195 clearly benign prompts
• 53 controversial prompts

Harmful-request metrics use the 92 clearly harmful prompts as their denominator. The benign overrefusal rate uses the 195 clearly benign prompts as its denominator.

The proxy set participated in candidate-model selection. Therefore, the reported safety results should be interpreted as same-protocol validation results rather than performance on an independent held-out safety test set.

The seven targeted prompts participated in targeted safety repair. Therefore, the 7/7 result is a regression-test result rather than an independent generalization result.

Safe, Unsafe, and Controversial are mutually exclusive response-safety labels. Explicit refusal is a separate behavioral label and is usually a subset of safe responses.

Usage

Install a recent version of Transformers:

pip install -U transformers accelerate safetensors

Load and run the model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "whleric/Qwen3-0.6B-SafeMath-V3"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Solve: If 3x + 5 = 20, what is x?",
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

inputs = tokenizer(
    text,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

generated_tokens = outputs[0, inputs["input_ids"].shape[1]:]

response = tokenizer.decode(
    generated_tokens,
    skip_special_tokens=True,
)

print(response)

The example uses non-thinking mode and deterministic decoding.

Intended Use

This model is intended for research and evaluation involving:

• safety alignment;
• harmful-request handling;
• mathematical reasoning;
• capability trade-offs in parameter-efficient fine-tuning;
• small-scale language-model experimentation.

The reported results are limited to the listed benchmarks and safety proxy set; performance on unseen prompts may vary.

Attribution

This model is derived from Qwen/Qwen3-0.6B.

The base model and this derivative release are distributed under the Apache License 2.0.