cognitivecomputations/dolphin-2.9.1-mixtral-1x22b converted to Mistral format. A Mixtral model with a single expert is mathematically equivalent to the corresponding Mistral model. This allows to remove 344k parameters and to avoid software bugs when encountering a Mixtral with 1 expert.

Note that ChatML is entirely broken in the original and converted model. I have no plausible explanation why it's broken. Alpaca seems to work even though the model is not trained on it.

Original model card below.

Dolphin 2.9.1 Mixtral 1x22b 🐬

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

Discord: https://discord.gg/cognitivecomputations

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

axolotl version: 0.4.0

base_model: cognitivecomputations/mixtral-1x22b-base
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# trust_remote_code: true

# load_in_8bit: true
# load_in_4bit: true
# strict: false

datasets:
  - path: /workspace/datasets/dolphin-2.9/dolphin201-sharegpt2.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

chat_template: chatml
dataset_prepared_path: yi34b-prepared
val_set_size: 0.01
output_dir: ./1x22b-out

# adapter: qlora
# lora_r: 16
# lora_alpha: 16
# lora_modules_to_save: [embed_tokens, lm_head]
# lora_dropout: 0.05
# lora_target_linear: true

# unfrozen_parameters:
# - ^lm_head.weight$
# - ^model.embed_tokens.weight$
# # input_layernorm layers
# - model.layers.0.input_layernorm
# - model.layers.1.input_layernorm
# - model.layers.2.input_layernorm
# - model.layers.3.input_layernorm
# - model.layers.4.input_layernorm
# - model.layers.5.input_layernorm
# - model.layers.6.input_layernorm
# - model.layers.7.input_layernorm
# - model.layers.8.input_layernorm
# - model.layers.9.input_layernorm
# - model.layers.10.input_layernorm
# - model.layers.11.input_layernorm
# - model.layers.12.input_layernorm
# - model.layers.13.input_layernorm
# - model.layers.14.input_layernorm
# - model.layers.15.input_layernorm
# - model.layers.16.input_layernorm
# - model.layers.17.input_layernorm
# - model.layers.18.input_layernorm
# - model.layers.19.input_layernorm
# - model.layers.20.input_layernorm
# - model.layers.21.input_layernorm
# - model.layers.22.input_layernorm
# - model.layers.23.input_layernorm
# # lm_head layers
# # mlp.down_proj layers
# - model.layers.17.mlp.down_proj
# - model.layers.18.mlp.down_proj
# - model.layers.19.mlp.down_proj
# - model.layers.20.mlp.down_proj
# - model.layers.21.mlp.down_proj
# - model.layers.22.mlp.down_proj
# - model.layers.23.mlp.down_proj
# - model.layers.24.mlp.down_proj
# - model.layers.25.mlp.down_proj
# - model.layers.26.mlp.down_proj
# - model.layers.27.mlp.down_proj
# - model.layers.28.mlp.down_proj
# - model.layers.29.mlp.down_proj
# - model.layers.30.mlp.down_proj
# - model.layers.31.mlp.down_proj
# - model.layers.32.mlp.down_proj
# - model.layers.33.mlp.down_proj
# - model.layers.34.mlp.down_proj
# - model.layers.35.mlp.down_proj
# - model.layers.36.mlp.down_proj
# - model.layers.37.mlp.down_proj
# - model.layers.38.mlp.down_proj
# - model.layers.39.mlp.down_proj
# - model.layers.40.mlp.down_proj
# # mlp.gate_proj layers
# - model.layers.51.mlp.gate_proj
# - model.layers.50.mlp.gate_proj
# - model.layers.53.mlp.gate_proj
# - model.layers.52.mlp.gate_proj
# - model.layers.49.mlp.gate_proj
# - model.layers.45.mlp.gate_proj
# - model.layers.46.mlp.gate_proj
# - model.layers.47.mlp.gate_proj
# - model.layers.57.mlp.gate_proj
# - model.layers.48.mlp.gate_proj
# - model.layers.56.mlp.gate_proj
# - model.layers.41.mlp.gate_proj
# - model.layers.54.mlp.gate_proj
# - model.layers.43.mlp.gate_proj
# - model.layers.44.mlp.gate_proj
# - model.layers.60.mlp.gate_proj
# - model.layers.55.mlp.gate_proj
# - model.layers.40.mlp.gate_proj
# - model.layers.42.mlp.gate_proj
# - model.layers.58.mlp.gate_proj
# - model.layers.36.mlp.gate_proj
# - model.layers.37.mlp.gate_proj
# - model.layers.38.mlp.gate_proj
# - model.layers.39.mlp.gate_proj
# # mlp.up_proj layers
# - model.layers.50.mlp.up_proj
# - model.layers.51.mlp.up_proj
# - model.layers.41.mlp.up_proj
# - model.layers.49.mlp.up_proj
# - model.layers.43.mlp.up_proj
# - model.layers.44.mlp.up_proj
# - model.layers.40.mlp.up_proj
# - model.layers.45.mlp.up_proj
# - model.layers.47.mlp.up_proj
# - model.layers.48.mlp.up_proj
# - model.layers.46.mlp.up_proj
# - model.layers.42.mlp.up_proj
# - model.layers.39.mlp.up_proj
# - model.layers.36.mlp.up_proj
# - model.layers.37.mlp.up_proj
# - model.layers.38.mlp.up_proj
# - model.layers.56.mlp.up_proj
# - model.layers.57.mlp.up_proj
# - model.layers.53.mlp.up_proj
# - model.layers.31.mlp.up_proj
# - model.layers.32.mlp.up_proj
# - model.layers.34.mlp.up_proj
# - model.layers.35.mlp.up_proj
# - model.layers.33.mlp.up_proj
# # model.embed_tokens layers
# # model.norm layers
# # post_attention_layernorm layers
# - model.layers.0.post_attention_layernorm
# - model.layers.1.post_attention_layernorm
# - model.layers.2.post_attention_layernorm
# - model.layers.3.post_attention_layernorm
# - model.layers.4.post_attention_layernorm
# - model.layers.5.post_attention_layernorm
# - model.layers.6.post_attention_layernorm
# - model.layers.7.post_attention_layernorm
# - model.layers.8.post_attention_layernorm
# - model.layers.9.post_attention_layernorm
# - model.layers.10.post_attention_layernorm
# - model.layers.11.post_attention_layernorm
# - model.layers.12.post_attention_layernorm
# - model.layers.13.post_attention_layernorm
# - model.layers.14.post_attention_layernorm
# - model.layers.15.post_attention_layernorm
# - model.layers.16.post_attention_layernorm
# - model.layers.17.post_attention_layernorm
# - model.layers.18.post_attention_layernorm
# - model.layers.19.post_attention_layernorm
# - model.layers.20.post_attention_layernorm
# - model.layers.21.post_attention_layernorm
# - model.layers.22.post_attention_layernorm
# - model.layers.23.post_attention_layernorm
# # self_attn.k_proj layers
# - model.layers.42.self_attn.k_proj
# - model.layers.41.self_attn.k_proj
# - model.layers.39.self_attn.k_proj
# - model.layers.35.self_attn.k_proj
# - model.layers.28.self_attn.k_proj
# - model.layers.79.self_attn.k_proj
# - model.layers.43.self_attn.k_proj
# - model.layers.32.self_attn.k_proj
# - model.layers.73.self_attn.k_proj
# - model.layers.31.self_attn.k_proj
# - model.layers.29.self_attn.k_proj
# - model.layers.76.self_attn.k_proj
# - model.layers.30.self_attn.k_proj
# - model.layers.40.self_attn.k_proj
# - model.layers.33.self_attn.k_proj
# - model.layers.78.self_attn.k_proj
# - model.layers.34.self_attn.k_proj
# - model.layers.37.self_attn.k_proj
# - model.layers.45.self_attn.k_proj
# - model.layers.44.self_attn.k_proj
# - model.layers.71.self_attn.k_proj
# - model.layers.26.self_attn.k_proj
# - model.layers.74.self_attn.k_proj
# - model.layers.27.self_attn.k_proj
# # self_attn.o_proj layers
# - model.layers.35.self_attn.o_proj
# - model.layers.34.self_attn.o_proj
# - model.layers.37.self_attn.o_proj
# - model.layers.33.self_attn.o_proj
# - model.layers.31.self_attn.o_proj
# - model.layers.27.self_attn.o_proj
# - model.layers.38.self_attn.o_proj
# - model.layers.24.self_attn.o_proj
# - model.layers.39.self_attn.o_proj
# - model.layers.43.self_attn.o_proj
# - model.layers.29.self_attn.o_proj
# - model.layers.0.self_attn.o_proj
# - model.layers.50.self_attn.o_proj
# - model.layers.32.self_attn.o_proj
# - model.layers.45.self_attn.o_proj
# - model.layers.30.self_attn.o_proj
# - model.layers.60.self_attn.o_proj
# - model.layers.23.self_attn.o_proj
# - model.layers.18.self_attn.o_proj
# - model.layers.67.self_attn.o_proj
# - model.layers.57.self_attn.o_proj
# - model.layers.20.self_attn.o_proj
# - model.layers.76.self_attn.o_proj
# - model.layers.28.self_attn.o_proj
# # self_attn.q_proj layers
# - model.layers.1.self_attn.q_proj
# - model.layers.6.self_attn.q_proj
# - model.layers.0.self_attn.q_proj
# - model.layers.5.self_attn.q_proj
# - model.layers.2.self_attn.q_proj
# - model.layers.7.self_attn.q_proj
# - model.layers.3.self_attn.q_proj
# - model.layers.4.self_attn.q_proj
# - model.layers.8.self_attn.q_proj
# - model.layers.9.self_attn.q_proj
# - model.layers.61.self_attn.q_proj
# - model.layers.10.self_attn.q_proj
# - model.layers.62.self_attn.q_proj
# - model.layers.36.self_attn.q_proj
# - model.layers.15.self_attn.q_proj
# - model.layers.11.self_attn.q_proj
# - model.layers.17.self_attn.q_proj
# - model.layers.60.self_attn.q_proj
# - model.layers.63.self_attn.q_proj
# - model.layers.64.self_attn.q_proj
# - model.layers.29.self_attn.q_proj
# - model.layers.30.self_attn.q_proj
# - model.layers.55.self_attn.q_proj
# - model.layers.34.self_attn.q_proj
# # self_attn.v_proj layers
# - model.layers.12.self_attn.v_proj
# - model.layers.16.self_attn.v_proj
# - model.layers.18.self_attn.v_proj
# - model.layers.19.self_attn.v_proj
# - model.layers.20.self_attn.v_proj
# - model.layers.21.self_attn.v_proj
# - model.layers.22.self_attn.v_proj
# - model.layers.23.self_attn.v_proj
# - model.layers.24.self_attn.v_proj
# - model.layers.25.self_attn.v_proj
# - model.layers.26.self_attn.v_proj
# - model.layers.27.self_attn.v_proj
# - model.layers.28.self_attn.v_proj
# - model.layers.29.self_attn.v_proj
# - model.layers.30.self_attn.v_proj
# - model.layers.31.self_attn.v_proj
# - model.layers.32.self_attn.v_proj
# - model.layers.33.self_attn.v_proj
# - model.layers.34.self_attn.v_proj
# - model.layers.35.self_attn.v_proj
# - model.layers.36.self_attn.v_proj
# - model.layers.37.self_attn.v_proj
# - model.layers.38.self_attn.v_proj
# - model.layers.39.self_attn.v_proj



sequence_len: 16384 
sample_packing: true
pad_to_sequence_len: true

# adapter: lora
# lora_model_dir:
# lora_r: 32
# lora_alpha: 16
# lora_dropout: 0.05
# lora_target_linear: true
# lora_fan_in_fan_out:

wandb_project: dolphin-mixtral1x22b
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-5

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint: /workspace/axolotl2/axolotl/1x22b-out/checkpoint-507
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 4
save_total_limit: 2
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
  bos_token: "<s>"
  # pad_token: "<unk>"
  unk_token: "<unk>"
tokens:
- "<|im_start|>"

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