You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
custom_chat_template is not compatible with qwen2.5-instruct series #63
Many thanks for sharing rstar series. I was wondering how to adjust custom_chat_template defined at rstar2_agent_trainer.yaml for RL training of qwen2.5-instruct.
For now, I get (AgentLoopWorker pid=3989193) ERROR:2025-11-06 13:45:39,760:Failed to decode tool call: Expecting ',' delimiter: line 3 column 1 (char 506)
Many thanks for sharing rstar series. I was wondering how to adjust custom_chat_template defined at rstar2_agent_trainer.yaml for RL training of qwen2.5-instruct.
For now, I get (AgentLoopWorker pid=3989193) ERROR:2025-11-06 13:45:39,760:Failed to decode tool call: Expecting ',' delimiter: line 3 column 1 (char 506)
Hello, I’d like to ask for some clarification.
I’m training Qwen3-14B-base on 8×H100, and during the early stage of training I encounter a large number of errors like:
Failed to decode tool call: Expecting property name enclosed in double quotes
Failed to decode tool call: Expecting ',' delimiter
Failed to decode tool call: Extra data
Error when executing tool: 'str' object has no attribute 'get'
Error when executing tool: 'WolframAlpha'
Error when executing tool: 'math'
I’m wondering whether this behavior is normal during the initial phase of training.
Activity
love-cfall commentedon Jan 26, 2026
Hello, I’d like to ask for some clarification.
I’m training Qwen3-14B-base on 8×H100, and during the early stage of training I encounter a large number of errors like:
Failed to decode tool call: Expecting property name enclosed in double quotes
Failed to decode tool call: Expecting ',' delimiter
Failed to decode tool call: Extra data
Error when executing tool: 'str' object has no attribute 'get'
Error when executing tool: 'WolframAlpha'
Error when executing tool: 'math'
I’m wondering whether this behavior is normal during the initial phase of training.