GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Terve0916
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Terve0916
Model Tree
Qwen/Qwen3-4B-Base
Input Modalities
Output Modalities
Supported Functionality
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen3-4B-Base |
| Training data | 3,200 examples from the GSM8K train split |
| On-policy rollouts | 12,800 |
| GRPO group size | 4 |
| LoRA | rank 16, alpha 32, dropout 0, all linear layers |
| Reward | Binary exact match on the final numeric answer |
| Learned reward model | None |
| Learning rate | 1e-5 with linear decay |
| Precision | BF16 |
| Selected checkpoint | Step 350, selected on a fixed validation split |
| Training seed | 42 |
| Hardware | One NVIDIA RTX 5090 32 GB |
The policy was trained for one epoch. The selected run used TRL 0.24.0, PEFT 0.17.1, Transformers 4.56.2, and vLLM 0.10.2.
Base and adapter evaluations used identical prompts, answer parsing, temperature, token budgets, and decoding seeds (42, 43, and 44).
| Benchmark | Base | LoRA-GRPO | Absolute gain |
|---|---|---|---|
| GSM8K official test (1,319 examples) | 84.69% | 88.91% | +4.22 pp |
| SVAMP OOD (1,000 examples) | 84.93% | 88.90% | +3.97 pp |
On GSM8K, the relative error reduction was 27.56%. Each paired decoding-seed comparison had a bootstrap 95% confidence interval above zero and an exact McNemar p-value below 0.001. SVAMP was not used for training or checkpoint selection, and all three decoding seeds improved.
The reported seeds are decoding seeds, not independent training runs. Only one GRPO training seed was run.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_model_id = "Qwen/Qwen3-4B-Base"adapter_id = "Terve0916/Qwen3-4B-Base-GRPO-GSM8K-LoRA"tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained(base_model_id,torch_dtype=torch.bfloat16,device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id)model.eval()
The training prompt asks the model to place its reasoning inside <think> tags
and its final numeric answer inside <answer> tags. Use the same prompting
protocol for results comparable to those reported above.
Training code, configurations, data construction, evaluation scripts, and
paired statistical analysis are available at
sheihui/qwen3-4b-lora-grpo.