Qwen3-14B AppWorld AgentEvolver checkpoints
This release contains model-weight exports from one AgentEvolver GRPO run on AppWorld. The run started from Qwen3-14B and completed 300 optimizer steps. The export set contains one checkpoint every 10 steps, from checkpoint-10 to checkpoint-300.
Each checkpoint includes roughly 59.1 GB of FP32 weights across two Safetensors shards, plus the model configuration, tokenizer, and chat template. The exports support inference with Transformers. Training state such as optimizer data is separate.
Download a checkpoint
hf download JunjieMeng/agentevolver-appworld-qwen3-14b-base-cot \
--include "checkpoint-300/*" --local-dir ./appworld-14b
Load ./appworld-14b/checkpoint-300 with Transformers. For comparable multi-turn AppWorld evaluations, preserve the sampled token history, including Qwen3 thinking tokens.