metadata

datasets:
  - Open-Orca/OpenOrca
language:
  - en
library_name: transformers
pipeline_tag: text-generation
license: apache-2.0

🐋 Mistral-7B-OpenOrca 🐋

OpenOrca - Mistral - 7B - 8k

We have used our own OpenOrca dataset to fine-tune on top of Mistral 7B. This dataset is our attempt to reproduce the dataset generated for Microsoft Research's Orca Paper. We use OpenChat packing, trained with Axolotl.

This release is trained on a curated filtered subset of most of our GPT-4 augmented data. It is the same subset of our data as was used in our OpenOrcaxOpenChat-Preview2-13B model.

HF Leaderboard evals place this model as #2 for all models smaller than 30B at release time, outperforming all but one 13B model.

This release provides a first: a fully open model with class-breaking performance, capable of running fully accelerated on even moderate consumer GPUs. Our thanks to the Mistral team for leading the way here.

We affectionately codename this model: "MistralOrca"

If you'd like to try the model now, we have it running on fast GPUs unquantized: https://huggingface.co/spaces/Open-Orca/Mistral-7B-OpenOrca

Want to visualize our full (pre-filtering) dataset? Check out our Nomic Atlas Map.

We are in-process with training more models, so keep a look out on our org for releases coming soon with exciting partners.

We will also give sneak-peak announcements on our Discord, which you can find here:

https://AlignmentLab.ai

or check the OpenAccess AI Collective Discord for more information about Axolotl trainer here:

https://discord.gg/5y8STgB3P3

Quantized Models

Quantized versions of this model are generously made available by TheBloke.

Prompt Template

We used OpenAI's Chat Markup Language (ChatML) format, with <|im_start|> and <|im_end|> tokens added to support this.

This means that, e.g., in oobabooga the "MPT-Chat" instruction template should work, as it also uses ChatML.

This formatting has also been set as a chat template, which means that lists of messages can be formatted for you with the apply_chat_template() method:

chat = [
  {"role": "user", "content": "Hello, how are you?"},
  {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
  {"role": "user", "content": "I'd like to show off how chat templating works!"},
]
tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

which will yield:

<|im_start|>user
Hello, how are you?<|im_end|>
<|im_start|>assistant
I'm doing great. How can I help you today?<|im_end|>
<|im_start|>user
I'd like to show off how chat templating works!<|im_end|>
<|im_start|>assistant

If you use tokenize=True and return_tensors="pt" instead, then you will get a tokenized and formatted conversation ready to pass to model.generate().

Example Prompt Exchange

<|im_start|>system
You are MistralOrca, a large language model trained by Alignment Lab AI. Write out your reasoning step-by-step to be sure you get the right answers!
<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
I am doing well!<|im_end|>
<|im_start|>user
Please tell me about how mistral winds have attracted super-orcas.<|im_end|>

Inference

See this notebook for inference details.

Note that you need the development snapshot of Transformers currently, as support for Mistral hasn't been released into PyPI yet:

pip install git+https://github.com/huggingface/transformers

Evaluation

HuggingFace Leaderboard Performance

We have evaluated using the methodology and tools for the HuggingFace Leaderboard, and find that we have dramatically improved upon the base model. We find 105% of the base model's performance on HF Leaderboard evals, averaging 65.33.

At release time, this beats all 7B models, and all but one 13B.

Metric	Value
MMLU (5-shot)	61.73
ARC (25-shot)	63.57
HellaSwag (10-shot)	83.79
TruthfulQA (0-shot)	52.24
Avg.	65.33

We use Language Model Evaluation Harness to run the benchmark tests above, using the same version as the HuggingFace LLM Leaderboard.

AGIEval Performance

We compare our results to the base Mistral-7B model (using LM Evaluation Harness).

We find 129% of the base model's performance on AGI Eval, averaging 0.397. As well, we significantly improve upon the official mistralai/Mistral-7B-Instruct-v0.1 finetuning, achieving 119% of their performance.

BigBench-Hard Performance

We find 119% of the base model's performance on BigBench-Hard, averaging 0.416.

GPT4ALL Leaderboard Performance

We gain a slight edge over our previous releases, again topping the leaderboard, averaging 72.38.

Dataset

We used a curated, filtered selection of most of the GPT-4 augmented data from our OpenOrca dataset, which aims to reproduce the Orca Research Paper dataset.

Training

We trained with 8x A6000 GPUs for 62 hours, completing 4 epochs of full fine tuning on our dataset in one training run. Commodity cost was ~$400.

Citation

@software{lian2023mistralorca1
  title = {MistralOrca: Mistral-7B Model Instruct-tuned on Filtered OpenOrcaV1 GPT-4 Dataset},
  author = {Wing Lian and Bleys Goodson and Guan Wang and Eugene Pentland and Austin Cook and Chanvichet Vong and "Teknium"},
  year = {2023},
  publisher = {HuggingFace},
  journal = {HuggingFace repository},
  howpublished = {\url{https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca},
}
@misc{mukherjee2023orca,
      title={Orca: Progressive Learning from Complex Explanation Traces of GPT-4}, 
      author={Subhabrata Mukherjee and Arindam Mitra and Ganesh Jawahar and Sahaj Agarwal and Hamid Palangi and Ahmed Awadallah},
      year={2023},
      eprint={2306.02707},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
@misc{longpre2023flan,
      title={The Flan Collection: Designing Data and Methods for Effective Instruction Tuning}, 
      author={Shayne Longpre and Le Hou and Tu Vu and Albert Webson and Hyung Won Chung and Yi Tay and Denny Zhou and Quoc V. Le and Barret Zoph and Jason Wei and Adam Roberts},
      year={2023},
      eprint={2301.13688},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}