Qwen3-1.7B-Math
This checkpoint is released for mathematical reasoning experiments in the G²RPO-A project. It was fine-tuned from Qwen/Qwen3-1.7B using the mathematical guidance data now available as G2RPO-A/Math-Curriculum-1K.
Model identity
The original repository name and generated model card incorrectly labeled this checkpoint as Qwen3-0.6B. The checkpoint's config.json matches the official Qwen3-1.7B configuration on the key architecture fields: 28 layers, hidden size 2048, intermediate size 6144, 16 attention heads, 8 key/value heads, and vocabulary size 151936. The model owner identifies it as the 1.7B mathematical-task checkpoint. This documentation correction does not modify the model weights or configuration.
Quick start
from transformers import pipeline
generator = pipeline(
"text-generation",
model="G2RPO-A/Qwen3-1.7B-Math",
device="cuda",
)
question = "Solve 2x + 3 = 11. Explain your reasoning and give the value of x."
output = generator(
[{"role": "user", "content": question}],
max_new_tokens=2048,
do_sample=True,
temperature=0.6,
top_p=0.95,
top_k=20,
return_full_text=False,
)[0]
print(output["generated_text"])
Training and framework versions
The original training-generated card records GRPO training with TRL and the following framework versions:
- TRL: 0.15.2
- Transformers: 4.52.3
- PyTorch: 2.5.1
- Datasets: 3.6.0
- Tokenizers: 0.21.1
Use Transformers 4.52.3 to match that record; the older 4.49.0 Open-R1 dependency pin does not support Qwen3 out of the box.
See the ACL 2026 paper for the project methodology. This card does not assign the paper's benchmark scores to this checkpoint without a matching evaluation record.