KaeriJenti/kaori-34b-v4

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Dec 22, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

KaeriJenti/kaori-34b-v4 is a 34 billion parameter language model fine-tuned by Kaeri and Jenti, built upon an unspecified base architecture. It was trained using a Supervised Fine-Tuning (SFT) strategy on a curated mix of Open-Platypus, Dolphin, and OpenOrca datasets. This model is specifically designed with careful data contamination filtering to avoid common benchmark datasets, making it suitable for general-purpose conversational AI and instruction-following tasks where benchmark integrity is crucial.

Loading preview...

KaeriJenti/kaori-34b-v4 Overview

KaeriJenti/kaori-34b-v4 is a 34 billion parameter language model developed through supervised fine-tuning (SFT) by Kaeri and Jenti. The model was trained using a LoRA finetuning type over 3 epochs on A100 GPUs.

Training Details

The model's training regimen focused on a specific blend of datasets:

  • Open-Platypus: 100% of this dataset was utilized.
  • Dolphin: 5% of this dataset was incorporated.
  • OpenOrca: 10% of this dataset was used.

Data Contamination Filtering

A key aspect of kaori-34b-v4's development involved rigorous data contamination filtering. The creators explicitly excluded GSM8k samples and applied similarity filtering against a list of common benchmark tasks, including cot_gsm8k, drop, winogrande, ai2_arc, and hellaswag. This careful approach aims to ensure the model's performance is not artificially inflated by exposure to benchmark data during training.

Use Cases

This model is well-suited for applications requiring a general-purpose instruction-following language model, particularly where the integrity of evaluation against standard benchmarks is a concern due to the deliberate filtering of common benchmark datasets from its training data.