Quick tests show it's good but not as sharp as the base model, using just some few shot prompts looking for precision when asking specifics about methods in a process. More tests will have to be done to compare this and WizardLM-7B to see how much the finetuning/new EOS did to Llama-3-8B.

Notably, cognitivecomputations uses a single EOS token. This fixes the garbled output bug. Hooray! It may however prevent some intended behavior of Llama3's internal monologue/thoughts that adds to the model's apparent sharpness. Download Meta's original weights and load manually in python to see what it's capable of as a comparison. We're all awaiting any fixes to llama.cpp and/or the base gguf structure. In the meantime this dolphin is a good fix and excellent work.

conversion notes: I converted the original safetensors to f32 to preserve the fidelity from bf16, then quantized ggufs from there. Not sure what most ggufs on hf are doing if they don't say.

size notes: Windows users, go for q3-k-s. FreeBSD users, you're the real heroes. Others, use the biggest one that works on your machine.

I just copied the original model card this time.

.-=~ Original Model Card ~=-.

Dolphin 2.9 Llama 3 8b 🐬

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

Discord: https://discord.gg/8fbBeC7ZGx

My appreciation for the sponsors of Dolphin 2.9:

Crusoe Cloud - provided excellent on-demand 10xL40S node

This model is based on Llama-3-8b, 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 2.5 days on 8x L40S provided by Crusoe Cloud

This model was trained FFT on all parameters, 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 has a variety of instruction, conversational, and coding skills. It also has initial agentic abilities and supports function calling.

Dolphin is uncensored. I 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. I 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.

See axolotl config

axolotl version: 0.4.0

yaml

base_model: meta-llama/Meta-Llama-3-8B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
tokenizer_use_fast: false


load_in_8bit: false
load_in_4bit: false
strict: false
model_config:

datasets:
  - path: /workspace/datasets/dolphin-2.9/dolphin201-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/Ultrachat200kunfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/dolphin-coder-translate-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/dolphin-coder-codegen-sharegpt2.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/m-a-p_Code-Feedback-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/m-a-p_CodeFeedback-Filtered-Instruction-sharegpt-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/not_samantha_norefusals.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/Orca-Math-resort-unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/agent_instruct_react_unfiltered.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_instruct_j1s1_3k_unfiltered.jsonl
    type: sharegpt  
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_negative_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_react_10p_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/toolbench_tflan_cot_30p_unfiltered.jsonl
    type: sharegpt
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/openhermes200k_unfiltered.jsonl
    type: sharegpt 
    conversation: chatml
  - path: /workspace/datasets/dolphin-2.9/SystemConversations.jsonl
    type: sharegpt
    conversation: chatml

chat_template: chatml
    

dataset_prepared_path: /workspace/datasets/dolphin-2.9/thingy
val_set_size: 0.0002
output_dir: ./out

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 4
micro_batch_size: 3
num_epochs: 3
logging_steps: 1
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

wandb_project: dolphin-2.9-mixtral-8x22b
wandb_watch:
wandb_run_id:
wandb_log_model:

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
saves_per_epoch: 4
save_total_limit: 2
save_steps:
evals_per_epoch: 4
eval_sample_packing: false
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|end_of_text|>"
tokens:
  - "<|im_start|>"
  - "<|im_end|>"

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 3
eval_batch_size: 3
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 4
total_train_batch_size: 96
total_eval_batch_size: 24
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 7
num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.146	0.0005	1	1.1064
0.6962	0.2501	555	0.6636
0.6857	0.5001	1110	0.6503
0.6592	0.7502	1665	0.6419
0.6465	1.0002	2220	0.6317
0.5295	1.2395	2775	0.6408
0.5302	1.4895	3330	0.6351
0.5188	1.7396	3885	0.6227
0.521	1.9896	4440	0.6168
0.3968	2.2289	4995	0.6646
0.3776	2.4789	5550	0.6619
0.3983	2.7290	6105	0.6602

Framework versions

Transformers 4.40.0
Pytorch 2.2.2+cu121
Datasets 2.18.0
Tokenizers 0.19.1

gobean/dolphin-2.9-llama3-8b.llamafile