Stability-AI / StableLM

StableLM: Stability AI Language Models
Apache License 2.0
15.84k stars 1.03k forks source link

Big difference between the before-cooldown-ckpt and the final checkpoint in the results of downstream tasks? #104

Open siqi13579 opened 7 months ago

siqi13579 commented 7 months ago

We tested the checkpoint before cooldown and the final checkpoint of stable lm2 1.6b , and found that there was a big difference in the results of downstream tasks. The final checkpoint significantly improved the results of downstream tasks. Are there any special strategies for the cooldown phase? arc(25shot) hellaswag(10shot) mmlu(5-shot) truthfulqa winnogrande(5-shot) gsm(5-shot) 43.52 70.3 39.8 36.61 64.17 17.29 38.4 67.59 30 34.9 61.96 7.35