Skip to content

🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.

[feat] repetition-penalty

jingyaogongfa6df82ff8b1
diff --git a/eval_model.py b/eval_model.py
index 1169865..e73b925 100755
--- a/eval_model.py
+++ b/eval_model.py
@@ -169,7 +169,8 @@ def main():
eos_token_id=tokenizer.eos_token_id,
streamer=streamer,
top_p=args.top_p,
- temperature=args.temperature
+ temperature=args.temperature,
+ repetition_penalty=1.0 # 重复惩罚力度,1.0使用默认分布
)
response = tokenizer.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)