Student model, after fine-tuning, improves upon the performance of the basemodel on two benchmarks: truthfulqa and gsm8k
-
truthfulqa: student = 39.29 vs base = 38.3
-
gsm8k: student = 17.06 vs base = 16.3
Benchmarks
aloobun/d-Qwen1.5-0.5B:
| Avg. | Arc | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|---|---|---|---|---|---|
| 38.07 | 30.29 | 47.75 | 38.21 | 39.29 | 55.8 | 17.06 |
Qwen/Qwen1.5-0.5B:
| Avg. | Arc | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|---|---|---|---|---|---|
| 38.62 | 31.48 | 49.05 | 39.35 | 38.3 | 57.22 | 16.3 |
I will train it longer in my next run; can do better.
- This is a distillation experiment with Qwen1.5-1.8B as teacher and Qwen1.5-0.5B as student model respectively.
- Samples were taken from the Pile dataset.
- optimizer: SM3, scheduler: cosine with warmup, lr=2e-5
Qwen1.5 is a language model series including decoder language models of different model sizes. For each size, we release the base language model and the aligned chat model. It is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, mixture of sliding window attention and full attention, etc. Additionally, we have an improved tokenizer adaptive to multiple natural languages and codes. For the beta version, temporarily we did not include GQA and the mixture of SWA and full attention.