See axolotl config

axolotl version: 0.4.0

base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: true
strict: false
chat_template: inst

datasets:
  - path: ./data/raw_format/tool_used_training.jsonl
    type: sharegpt
    conversation: mistral
  - path: ./data/raw_format/tool_not_used_training.jsonl
    type: sharegpt
    conversation: mistral
  - path: ./data/raw_format/no_tools_training.jsonl
    type: sharegpt
    conversation: mistral
  - path: ./data/akoksal_lon_form.jsonl
    type: sharegpt
    conversation: mistral
  - path: ./data/dolly.jsonl
    type: sharegpt
    conversation: mistral

dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./mixtral-lora-2-epochs-r64

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_fan_in_fan_out:
hub_model_id: liuylhf/mixtral-lora
hub_strategy: end
# lora_target_linear: true
model_config:
  output_router_logits: true
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: function-call
wandb_name: mixtral-instruct-raw-data-v3
wandb_log_model: end

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.001
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

# loss_watchdog_threshold: 5.0
# loss_watchdog_patience: 3

warmup_steps: 10
# evals_per_epoch: 20
eval_steps: 0.05
save_steps: 0.1
eval_table_size:
eval_max_new_tokens: 256
# saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0

mixtral-lora

This model is a fine-tuned version of mistralai/Mixtral-8x7B-Instruct-v0.1 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4469

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: 0.001
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
3.3559	0.0	1	3.6627
0.548	0.1	43	0.5124
0.2747	0.2	86	0.4845
0.4202	0.31	129	0.4760
0.4662	0.41	172	0.4690
0.4605	0.51	215	0.4640
0.2909	0.61	258	0.4620
0.3941	0.71	301	0.4600
0.4185	0.82	344	0.4573
0.395	0.92	387	0.4558
0.2725	1.0	430	0.4534
0.2789	1.1	473	0.4525
0.4126	1.21	516	0.4511
0.3277	1.31	559	0.4506
0.3591	1.41	602	0.4493
0.3665	1.51	645	0.4487
0.5551	1.62	688	0.4471
0.3363	1.72	731	0.4473
0.3117	1.82	774	0.4471
0.496	1.92	817	0.4469

Framework versions

• PEFT 0.8.2
• Transformers 4.39.0.dev0
• Pytorch 2.2.0+cu121
• Datasets 2.17.1
• Tokenizers 0.15.0