Dolphin 2.9.1 Llama 3 70b 🐬

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

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

👁 Image

We have retrained our LLama-3-70b fine tune to address behavioral issues in the initial 2.9 dataset. Specifically, Systemchat was causing the model to be too reliant on the system prompt. Additionally, it had an occasional quirk that would cause the model to overly reference the system prompt. We also found generation length was at times not sufficient for any given task. We identified the culprit as Ultrachat. Accounting for these concerns, we removed systemchat and ultrachat from the dataset. It is otherwise identical to dolphin-2.9.

Our appreciation for the sponsors of Dolphin 2.9.1:

• Crusoe Cloud - provided excellent on-demand 8xL40S node
• OnDemand - provided inference sponsorship

This model is based on Llama-3-70b, and is governed by META LLAMA 3 COMMUNITY LICENSE AGREEMENT

The base model has 8k context, and the full-weight fine-tuning was with 4k sequence length.

It took 3 days on an 8x H100 provided by Crusoe Cloud

This model was trained FFT on parameters selected by Laser Scanner, using ChatML prompt template format.

example:

<|im_start|>system
You are Dolphin, a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Dolphin-2.9.1 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 according to Meta's Llama license. We grant permission for any use, including commercial, that falls within accordance with Meta's Llama-3 license. Dolphin was trained on data generated from GPT4, among other models.

Evals

👁 image/png

Training

👁 Built with Axolotl

workspace/axolotl/llama-70b

This model is a fine-tuned version of meta-llama/Meta-Llama-3-70B on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4808

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: 5
• num_epochs: 3

Training results

Framework versions

• Transformers 4.40.2
• Pytorch 2.2.2+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1

Base model

meta-llama/Meta-Llama-3-70B

(47)