e2h-polaris-qwen3-4b-full-mixed
a Qwen3-4B checkpoint trained on math problems
1 min readupdated
This is a 4-billion-parameter Qwen/Qwen3-4B language model fine-tuned with GRPO (Group Relative Policy Optimization) on the Polaris-Dataset-53K, split into easy and hard math problems. Training relied on rule-based verification for simple examples and majority voting for harder ones. This checkpoint captures step 50 out of 1,200 planned training steps and targets stronger mathematical reasoning on top of the base architecture. It is useful for researchers and developers working to improve logic and math performance in compact language models.
Key features
- Base model: Qwen/Qwen3-4B with 4.41B parameters.
- Trained on 26,790 problems from the Polaris dataset (a mix of easy and hard tasks).
- Uses GRPO with a hybrid reward setup (answer verification and majority voting).
- Weights are saved in bf16 format and merged from verl FSDP shards.
What you can do
- Run the model to solve complex math problems and logic tasks.
- Benchmark chain-of-thought reasoning quality against the original Qwen3-4B.
- Plug the model into inference pipelines using vLLM or standard transformers.
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "davidafrica/e2h-polaris-qwen3-4b-full-mixed"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
Read more
DetailsTraining method: verl 0.4.1 (GRPO)
Response length: up to 8,192 tokens
Training step: 50
Weight precision: bf16










Comments