Dolphin 2.9.1 Mixtral 1x22b 🐬

Curated and trained by Eric Hartford, Lucas Atkins, and Fernando Fernandes, and Cognitive Computations

👁 Discord
Discord: https://discord.gg/cognitivecomputations

👁 Image

This model is based on Dolphin-2.9-Mixtral-8x22b, and is Apache-2.0 licensed.

The base model has 64k context, and the full-weight fine-tuning was with 16k sequence length.

It took 27 hours on 8xH100 provided by Crusoe Cloud.

This model was fully fine-tuned, targeting all layers.

The model is an extracted expert using SLERP and a custom script that we've open-sourced. It extracts a single expert which is the combined SLERP of all 8 experts from a Mixtral architecture. We decided to not fully convert to a dense model, for the sake of trying to keep as much of the original model's performance as possible, as this process is already quite surgical and there are a lot of variables to take into account.

Dolphin-2.9 has a variety of instruction, conversational, and coding skills. It also has initial agentic abilities and supports function calling.

Dolphin is uncensored. We have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignment layer before exposing the model as a service. It will be highly compliant with any requests, even unethical ones. Please read my blog post about uncensored models. https://erichartford.com/uncensored-models You are responsible for any content you create using this model. Enjoy responsibly.

Dolphin is licensed under Apache 2.0. We grant permission for any use, including commercial, as long as it complies with the Apache-2.0 license. Dolphin was trained using data generated from GPT-4, among other models. For more details on the extraction process of the expert model, visit our GitHub repository: https://github.com/cognitivecomputations/extract-expert/tree/main

Evals

👁 image/png

👁 Built with Axolotl

1x22b-out

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Framework versions

• Transformers 4.40.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1

Base model

mistral-community/Mixtral-8x22B-v0.1

(12)