Trained on cleaned datasets.
Safetensors: mondk/MiniCPM5-1B-MSH-SFT-SAFETENSORS
Datasets:
- mondk/UltraData-SFT-2605-CLEANED
- mondk/Greetings-hi-for-train-Msh-v2
Improvements:
- Reduced Chinese language knowledge (prevents the model from switching to another language while conversing in English).
Limitations:
- Model may be faster due to its lighter weight, but may not be as thoroughly trained as OpenBMB
benchmark_here
Based on my benchmarks (results may vary):
| mondk/MiniCPM5-1B-MSH-SFT | Qwen/Qwen3.5-4B | |
|---|---|---|
| Pass rate | 27% | 10% |
| Easy | 2/15 | 2/15 |
| Medium | 9/25 | 2/25 |
| Hard | 5/20 | 2/20 |
| Latency/question | 14.8s | 1.8s |
For maximum performance, set temp = 0.8
ty