[Feature]: Add conversion script from training checkpoints to final Hugging Face format #29
Description
Hi, thanks for open-sourcing AgentEvolver, it’s been very helpful for my experiments.
I’d like to request a feature related to model export / conversion.
Motivation
Right now, after training with AgentEvolver, I end up with checkpoints that are not directly usable by some inference frameworks that expect a standard Hugging Face model structure (e.g., for use with vLLM).
To run the trained model with vLLM, I currently have to:
Manually download the model_merger or related conversion scripts from the verl repository (or other external sources).
Run those scripts myself to merge/convert the checkpoints into a final Hugging Face-compatible directory (with config.json, tokenizer, and the consolidated pytorch_model-*.bin or model.safetensors weights).
This extra manual step is a bit inconvenient and error-prone, especially when trying to integrate AgentEvolver into a reproducible pipeline.
Feature Request
It would be very helpful if AgentEvolver could include:
A built-in script or CLI command to convert / merge the training checkpoints into a final Hugging Face-style model directory, ready to be used by vLLM and other inference frameworks.