GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zbeeb
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zbeeb
Model Tree
Qwen/Qwen2.5-0.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Fine-tuned from Qwen/Qwen2.5-0.5B-Instruct on a 100K math reasoning SFT mixture for 2 epochs with learning rate 1e-5.
Prompt format used during training:
text
System: Please reason step by step, and put your final answer within \boxed{}.User: {problem}Assistant: {solution}Training mixture:
| Source | Count |
|---|---|
| nvidia/OpenMathReasoning CoT | 40,000 |
| AI-MO/NuminaMath-1.5 filtered, no AMC/AIME source | 25,000 |
| meta-math/MetaMathQA | 15,000 |
| MATH train, especially levels 4-5 | 15,000 |
| GSM8K train | 5,000 |
Training summary:
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen2.5-0.5B-Instruct |
| Training examples | 100,000 |
| Epochs | 2 |
| Learning rate | 1e-5 |
| Max sequence length | 4096 |
| Effective batch size | 32 |
| Final train loss | 0.6923 |
| Final token accuracy | about 0.803 |
Evaluation results are not included in this model card yet.