Qwen3 8B AppWorld AgentEvolver checkpoints
This repository contains 30 full-parameter checkpoints from one AppWorld reinforcement-learning run. The run initialized from the original Qwen3-8B weights, not from an earlier SFT or RL checkpoint. It completed 300 optimizer steps using AgentEvolver. Checkpoints were saved every 10 steps, from checkpoint-10 through checkpoint-300.
Each directory is a standalone Transformers model with weights, configuration, chat template, and tokenizer. The weights are FP32 and approximately 32.8 GB per checkpoint. Optimizer state is not included in this repository.
To download one checkpoint:
hf download JunjieMeng/agentevolver-appworld-qwen3-8b-base-cot \
--include "checkpoint-100/*" --local-dir ./appworld-8b
Then load the local ./appworld-8b/checkpoint-100 directory with Transformers. Omit --include to download all 30 checkpoints.
Training and evaluation
- Framework: AgentEvolver, starting from upstream commit
a5a8db8689d6493b028107dcb4c27415441581dc. - Environment: AppWorld; 90 training tasks and 57 dev tasks.
- Algorithm: full-parameter GRPO, 32 tasks per batch and 8 sampled trajectories per task, with one optimizer update per batch.
- Learning rate:
1e-6; KL coefficient:0.001; sampling temperature:0.9; maximum 30 environment actions per trajectory. - Multi-turn training and evaluation preserved the exact sampled token history, including Qwen3 thinking tokens. Environment actions excluded the thinking text. For comparable multi-turn AppWorld evaluation, use a token-history-preserving inference path.
At step 300, the run logged AppWorld dev mean@8 = 0.6535 and best@8 = 0.8743. These are the training run's recorded environment-reward metrics over the 57 dev tasks, with 8 sampled trajectories per task.