Jeffrey Quesnelle
Jeffrey Quesnelle
resolves #415 This creates the "🥤SODA Synthetic Dialogue" dataset following the instructions as described in [datasets.md](https://github.com/LAION-AI/Open-Assistant/blob/main/docs/docs/data/datasets.md). Since this is the first dataset added to the repo, I followed the instructions...
When doing non-quantized LoRA, the dense trainable layers are instantiated without `dtype` and `device`, meaning that they will sit on the CPU. This can significantly slow down training and inference,...
This adds the `torch_dtype` and `load_in_8bit` quantization parameters for `transformers` models that allow loading larger models with lower VRAM requirements. The Llama and MPT subclasses load the model directly, so...
This is an updated version of #1606 that accounts for changes to the code since it was originally submitted. Adds support for llama.cpp's GBNF grammars, which enable very specific steering...
# What does this PR do? This adds the [YaRN RoPE scaling method](https://arxiv.org/abs/2309.00071) to the LLaMA-class of models. It can be activated for finetuned models by setting `rope_scaling.type = 'yarn'`...
This enables passing `readData` as an argument to `OpenDialogOptions` to enable reading selected files on Android. While the Dialog plugin code for Android correctly deals with `readData` and populates `base64Data`...
This fixes #345 by checking for all three combinations of bos/eos existence and using an appropriate template for each case
When using `DocumentTokenizer` if an `eos_token` is specified the tokenizer post processor is replaced with one that appends and EOS. However, this has the effect of NOT placing a BOS...