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.

train_agent: match numeric gt by value, not by substring

FeiyangPe888e12dabb8
diff --git a/trainer/train_agent.py b/trainer/train_agent.py
index c39a60e..b92075b 100644
--- a/trainer/train_agent.py
+++ b/trainer/train_agent.py
@@ -183,9 +183,10 @@ def rollout_batch(rollout_engine, tokenizer, messages_batch, tools_batch, num_ge
# ======== Reward 计算 ========
def validate_gt_in_text(text, gt_list):
+ # 数值型 gt 只按数值比对(子串会让 gt=9 命中 "19"、gt=100 命中 "1000");数字边界只看 ASCII 字母数字,中文紧挨数字("答案是59。")也能取到
text, text_num = str(text), str(text).replace(',', '')
- nums = [float(x) for x in re.findall(r'(?<![\w.])[-+]?\d+(?:\.\d+)?(?![\w.])', text_num)]
- return {g for g in gt_list if ((s := str(g).strip()) and s.lower() in text.lower()) or (re.fullmatch(r'[-+]?\d+(?:\.\d+)?', str(g).strip().replace(',', '')) and any(abs(float(str(g).strip().replace(',', '')) - n) < 1e-6 for n in nums))}
+ nums = [float(x) for x in re.findall(r'(?<![A-Za-z0-9_.])[-+]?\d+(?:\.\d+)?(?![A-Za-z0-9_]|\.\d)', text_num)]
+ return {g for g in gt_list if (s := str(g).strip()) and (any(abs(float(s.replace(',', '')) - n) < 1e-6 for n in nums) if re.fullmatch(r'[-+]?\d+(?:\.\d+)?', s.replace(',', '')) else s.lower() in text.lower())}
def calculate_rewards(prompts, completions, gt_batch, tools_batch, num_gen, reward_model=None, device="cuda", turn_outputs_batch=None, unfinished_batch=None):
rewards = torch.zeros(len(completions), device=device)