Dolphin 2.8 Mistral 7b v0.2 🐬
By Eric Hartford and Cognitive Computations
My appreciation for the sponsors of Dolphin 2.8:
- Crusoe Cloud - provided excellent on-demand 10xL40S node
- Winston Sou - Along with a generous anonymous sponsor, donated a massive personally owned compute resource!
- Abacus AI - my employer and partner in many things.
This model is based on Mistral-7b-v0.2 a new base model released by MistralAI on March 23, 2024 but they have not yet published on HuggingFace. Thanks to @alpindale for converting / publishing.
The base model has 32k context, and the full-weights fine-tune was with 16k sequence lengths.
It took 3 days on 10x L40S provided by Crusoe Cloud
Dolphin-2.8 has a variety of instruction, conversational, and coding skills.
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 to 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 Apache 2.0. I grant permission for any use including commercial. Dolphin was trained on data generated from GPT4 among other models.
Evals
{
"arc_challenge": {
"acc,none": 0.5921501706484642,
"acc_stderr,none": 0.014361097288449701,
"acc_norm,none": 0.6339590443686007,
"acc_norm_stderr,none": 0.014077223108470139
},
"gsm8k": {
"exact_match,strict-match": 0.4783927217589083,
"exact_match_stderr,strict-match": 0.013759618667051773,
"exact_match,flexible-extract": 0.5367702805155421,
"exact_match_stderr,flexible-extract": 0.013735191956468648
},
"hellaswag": {
"acc,none": 0.6389165504879506,
"acc_stderr,none": 0.004793330525656218,
"acc_norm,none": 0.8338976299541924,
"acc_norm_stderr,none": 0.00371411888431746
},
"mmlu": {
"acc,none": 0.6122347243982339,
"acc_stderr,none": 0.003893774654142997
},
"truthfulqa_mc2": {
"acc,none": 0.5189872652778472,
"acc_stderr,none": 0.014901128316426086
},
"winogrande": {
"acc,none": 0.7971586424625099,
"acc_stderr,none": 0.011301439925936643
}
}
axolotl version: 0.4.0
base_model: alpindale/Mistral-7B-v0.2-hf
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: /workspace/datasets/dolphin201-sharegpt2.jsonl
type: sharegpt
- path: /workspace/datasets/dolphin-coder-translate-sharegpt2.jsonl
type: sharegpt
- path: /workspace/datasets/dolphin-coder-codegen-sharegpt2.jsonl
type: sharegpt
- path: /workspace/datasets/m-a-p_Code-Feedback-sharegpt.jsonl
type: sharegpt
- path: /workspace/datasets/m-a-p_CodeFeedback-Filtered-Instruction-sharegpt.jsonl
type: sharegpt
- path: /workspace/datasets/not_samantha_norefusals.jsonl
type: sharegpt
- path: /workspace/datasets/openhermes2_5-sharegpt.jsonl
type: sharegpt
chat_template: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.001
output_dir: /workspace/dolphin-2.8-mistral-7b
sequence_len: 16384
sample_packing: true
pad_to_sequence_len: true
wandb_project: dolphin
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 8
micro_batch_size: 3
num_epochs: 4
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.000005
optimizer: adamw_bnb_8bit
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 10
eval_steps: 73
eval_table_size:
eval_table_max_new_tokens:
eval_sample_packing: false
saves_per_epoch:
save_steps: 73
save_total_limit: 2
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
eos_token: "<|im_end|>"
tokens:
- "<|im_start|>"
workspace/dolphin-2.8-mistral-7b
This model is a fine-tuned version of alpindale/Mistral-7B-v0.2-hf on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.4828
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: 5e-06
- train_batch_size: 3
- eval_batch_size: 3
- seed: 42
- distributed_type: multi-GPU
- num_devices: 10
- gradient_accumulation_steps: 8
- total_train_batch_size: 240
- total_eval_batch_size: 30
- optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 10
- num_epochs: 4
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 1.1736 | 0.0 | 1 | 1.0338 |
| 0.6106 | 0.36 | 73 | 0.5439 |
| 0.5766 | 0.72 | 146 | 0.5171 |
| 0.5395 | 1.06 | 219 | 0.5045 |
| 0.5218 | 1.42 | 292 | 0.4976 |
| 0.5336 | 1.78 | 365 | 0.4915 |
| 0.5018 | 2.13 | 438 | 0.4885 |
| 0.5113 | 2.48 | 511 | 0.4856 |
| 0.5066 | 2.84 | 584 | 0.4838 |
| 0.4967 | 3.19 | 657 | 0.4834 |
| 0.4956 | 3.55 | 730 | 0.4830 |
| 0.5026 | 3.9 | 803 | 0.4828 |
Framework versions
- Transformers 4.40.0.dev0
- Pytorch 2.2.1+cu121
- Datasets 2.18.0
- Tokenizers 0.15.0