OpenCerebrum-1.0-7B-DPO

OpenCerebrum-1.0-7B-DPO is an open-source language model fine-tuned from the alpindale/Mistral-7B-v0.2-hf base model on a diverse dataset aimed at replicating capabilities of Aether Research's proprietary Cerebrum model.

The model was fine-tuned on approximately 21,000 examples across 6 datasets spanning coding, math, science, reasoning, and general instruction-following. The goal was to assemble public datasets that could help the model achieve strong performance on benchmarks where Cerebrum excels.

I used the ChatML prompt format to train this model.

Model Details

• Base Model: alpindale/Mistral-7B-v0.2-hf
• Parameters: 7 billion
• Fine-Tuning Dataset Size: ~21,000 examples
• Fine-Tuning Data: Amalgamation of 6 public datasets
• Language: English
• License: Apache 2.0

Quants

• ExLlamaV2: https://huggingface.co/bartowski/OpenCerebrum-1.0-7b-DPO-exl2
• GGUF: https://huggingface.co/bartowski/OpenCerebrum-1.0-7b-DPO-GGUF
• AWQ: https://huggingface.co/solidrust/OpenCerebrum-1.0-7b-DPO-AWQ

Intended Use

OpenCerebrum-1.0-7B-DPO is intended to be a powerful open-source model for coding, math, science, and general question-answering and text generation tasks. Its diverse fine-tuning data aims to equip it with broad knowledge and reasoning capabilities.

However, as an open-source replica trained on a subset of data compared to the original Cerebrum, it may not match Cerebrum's full performance. Additionally, biases and limitations of the fine-tuning data may be reflected in the model's outputs.

Limitations and Biases

• The model may have biases and limitations inherited from its fine-tuning datasets. Thorough testing is needed to characterize these.
• With 21,000 training examples, the fine-tuning data is still limited compared to the proprietary Cerebrum data.
• As the model is based on a 7B parameter model, it has computational and memory constraints compared to larger models.

Training Details

The model was fine-tuned on the 6 datasets listed in the Datasets section, totaling approximately 21,000 examples. In the future, the fine-tuning dataset may be condensed to more closely match the ~500 example dataset reputedly used for the original Cerebrum model.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.78
AI2 Reasoning Challenge (25-Shot)	62.71
HellaSwag (10-Shot)	84.33
MMLU (5-Shot)	62.59
TruthfulQA (0-shot)	44.91
Winogrande (5-shot)	80.11
GSM8k (5-shot)	42.00