Instruction Pre-Training: Language Models are Supervised Multitask Learners
Paper • 2406.14491 • Published • 96
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("NewEden-Forge/grimjim_Kitsunebi-v1-Gemma2-8k-9B-exl2", dtype="auto")Quick Links
exl2 quant (measurement.json in main branch)
check revisions for quants
Kitsunebi-v1-Gemma2-8k-9B
This repo contains a merge of pre-trained Gemma 2 9B Instruct language models created using mergekit.
None of the components of this merge were trained for roleplay nor intended for it. Despite this, the resulting model can be used effectively for that function. The virtue of this model lies in its coherence, as opposed to textual richness.
This project utilizes HODACHI/EZO-Common-9B-gemma-2-it, a model based on gemma-2 and fine-tuned by Axcxept co., ltd. Its primary goal was to perform well in Japanese language tasks. Model training leveraged context-based synthesized instruction pre-training data for supervised multitask pre-training (abstract).
We also used princeton-nlp/gemma-2-9b-it-SimPO, a demonstration of Simple Preference Optimization (abstract).
Merge Details
Merge Method
This model was merged using the SLERP merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
slices:
- sources:
- model: princeton-nlp/gemma-2-9b-it-SimPO
layer_range: [0, 42]
- model: HODACHI/EZO-Common-9B-gemma-2-it
layer_range: [0, 42]
merge_method: slerp
base_model: HODACHI/EZO-Common-9B-gemma-2-it
parameters:
t:
- filter: self_attn
value: [0, 0.5, 0.3, 0.7, 1]
- filter: mlp
value: [1, 0.5, 0.7, 0.3, 0]
- value: 0.5
dtype: bfloat16
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="NewEden-Forge/grimjim_Kitsunebi-v1-Gemma2-8k-9B-exl2")