Edward Beeching

The functions for adding custom gltf to the model in gltf_export.py contain two unused arguments: buffer_data: ByteString, buffer_id: int = 0 These should be cleaned up, if they are not...

The user should be able to chose a GLTF file, load it and interact with it without the need for the python server.

Features for 0.9 release

3

My ideas for the 0.9 release are the following: - [x] Deprecation of using `gdrl` to launch training, I think it is better to have users refer to and use...

SFT_vlm script is missing the chat template

1

We did not include an option to add a chat template in the `sft_vlm.py` example script.

[BUG] MATH eval is calculated with non-greedy sampling

5

## Describe the bug Two metrics are reported for MATH, `Metrics.quasi_exact_match_math` and `Metrics.maj_at_4_math`, the first metric should be calculated using greedy generation, whereas the second with sampling and majority voting....

[Feature]: Expose option to load new model weights from disk

2

### 🚀 The feature, motivation and pitch In an async RL setting, we often want to perform fast generation with a vllm endpoint on a separate node and occasionally sync...