Clara Pohland
Clara Pohland
fixes #1447 - use `nanmean()` instead of `mean()` for losses to avoid nan losses - remove obsolete `accelerator.gather` for metrics as the metrics are all collected to cpu anyway and...
Within the training with KTO Trainer I occasionally experience `nan` values as rewards. I am running the training as a job on Ms Azure with one GPU (NVIDIA A100 80GB...
For DPOTrainer there exists the option to load the Adapter from SFT training twice, as in [Reference model considerations with PEFT - load-the-adapter-twice](https://huggingface.co/docs/trl/main/en/dpo_trainer#using-option-3---load-the-adapter-twice): ``` python # Initialize the trainer, without...
I recently experimented quite a bit with the BCO loss type in the KTO Trainer. This PR includes some changes that helped me to successfully and effectively run BCO on...
This should enable Upstream feature (https://docs.fluentbit.io/manual/configuration/upstream_servers) support for the Elasticsearch Output Plugin. It was tested in a local setup with two Elasticsearch instances (es-1 & es-2). Here is some example...