|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
pipeline_tag: text-generation |
|
|
datasets: |
|
|
- Delta-Vector/Hydrus-Preview-Tulu-3-SFT-Mix |
|
|
base_model: |
|
|
- arcee-ai/GLM-4-32B-Base-32K |
|
|
library_name: transformers |
|
|
tags: |
|
|
- instruct |
|
|
- code |
|
|
- chemistry |
|
|
- GLM |
|
|
--- |
|
|
|
|
|
<div align="center"> |
|
|
<img src="https://cdn-uploads.huggingface.co/production/uploads/66c26b6fb01b19d8c3c2467b/j9LpDp1Wup-m-IJ_XMh23.png" alt="Image" width="400" /> |
|
|
<br> |
|
|
<small><em>Promise I will never go blonde like Kanye</em></small> |
|
|
</div> |
|
|
|
|
|
--- |
|
|
# Overview |
|
|
|
|
|
Didn't really have any cool README ideas for this so we're just going with just whatever song i'm listening to rn and it happened to be `Baby i'm bleeding` |
|
|
|
|
|
Nevertheless, This is a finetune from the 32K context extended (or fixed?) Arcee GLM4 base - Trained shrimply with just the Tulu-SFT-Mixture *but* I removed Safety alignment examples. Came out pretty well, It uses chatML due to the GLM4 Format giving me a headache. It's a decently competant assistant although I haven't done any testing on how well the model performs at longer-contexts, nor have i done any RL afterwards to fix up it's edges. |
|
|
|
|
|
Think it should be a decent base for any future finetunes, I felt that GLM4 really wasn't given the proper time of day and it's a way better base then any Qwen3 model. |
|
|
|
|
|
# Quants |
|
|
|
|
|
GGUF: https://huggingface.co/mradermacher/GLM-Tulu-ChatML-GGUF |
|
|
|
|
|
Imatrix GGUF: https://huggingface.co/mradermacher/GLM-Tulu-ChatML-i1-GGUF |
|
|
|
|
|
# Prompting |
|
|
|
|
|
The model was trained with ChatML formatting |
|
|
|
|
|
``` |
|
|
"""<|im_start|>system |
|
|
system prompt<|im_end|> |
|
|
<|im_start|>user |
|
|
Hi there!<|im_end|> |
|
|
<|im_start|>assistant |
|
|
Nice to meet you!<|im_end|> |
|
|
<|im_start|>user |
|
|
Can I ask a question?<|im_end|> |
|
|
<|im_start|>assistant |
|
|
""" |
|
|
``` |
|
|
|
|
|
# Configs |
|
|
|
|
|
WandB : https://wandb.ai/new-eden/Training-A100/runs/05kktve8?nw=nwuserdeltavector |
|
|
|
|
|
This train took 15 hours on 8xB200s provided by Deepinfra and Cognitive Computations, Config is linked in the WandB |
|
|
|
|
|
# Credits |
|
|
|
|
|
Thank you to Lucy, Auri, NyxKrage, Creators of the Tulu-SFT-Mix and everyone at Anthracite & Allura |
|
|
|
|
|
|
