🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[fix] keep malformed tool calls from crashing agent RL training
Yi-111-adcbd31bf3d64
diff --git a/trainer/train_agent.py b/trainer/train_agent.pyindex 4619dbb..1b4ff9f 100644--- a/trainer/train_agent.py+++ b/trainer/train_agent.py@@ -76,8 +76,9 @@ CHECK_ARGS = {def parse_tool_calls(text):calls = []for m in re.findall(r'<tool_call>(.*?)</tool_call>', text, re.DOTALL):- try: calls.append(json.loads(m.strip()))- except: pass+ try: call = json.loads(m.strip())+ except: continue+ if isinstance(call, dict): calls.append(call) # 模型偶尔输出非对象JSON,直接丢弃避免下游崩溃return callsdef execute_tool(name, args):@@ -132,6 +133,7 @@ def rollout_single(rollout_engine, tokenizer, messages, tools, max_turns=3, max_if isinstance(raw, str):try: raw = json.loads(raw)except: raw = {}+ if not isinstance(raw, dict): raw = {} # arguments 结构异常按空参处理result = execute_tool(name, raw)result_str = (json.dumps(result, ensure_ascii=False) if result else '{"error": "tool not found"}')[:2048] # 防止天文数字撑爆tokenizermessages.append({"role": "tool", "content": result_str})@@ -225,6 +227,7 @@ def calculate_rewards(prompts, completions, gt_batch, tools_batch, num_gen, rewaif isinstance(raw, str):try: raw = json.loads(raw)except: raw = {}+ if not isinstance(raw, dict): raw = {} # arguments 结构异常按空参处理check = CHECK_ARGS.get(name)valid_call_count += int(bool(name in valid_names and check and check(raw)))tool_gap = abs(valid_call_count - len(gt)) + max(0, len(tool_calls) - valid_call_count) # tool数差值