liuylhf/mistral-qlora-v3

See axolotl config

axolotl version: 0.4.0

# base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
base_model: mistralai/Mistral-7B-Instruct-v0.2
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: ./data/tool_used_training.jsonl
    type: sharegpt
  - path: ./data/tool_not_used_training.jsonl
    type: sharegpt
  - path: ./data/no_tools_training.jsonl
    type: sharegpt

dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./mistral-instruct-qlora-v3

adapter: qlora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
# lora_target_modules:
#   - gate_proj
#   - down_proj
#   - up_proj
#   - q_proj
#   - v_proj
#   - k_proj
#   - o_proj

wandb_project: function-call
wandb_name: mixtral-v1
wandb_log_model: end

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
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: 0.1
save_steps: 0.1
eval_table_size:
eval_max_new_tokens: 256
# saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 1.0
fsdp:
fsdp_config:

mistral-instruct-qlora-v3

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.2 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3293

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: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.9129	0.02	1	0.9635
0.4847	0.11	7	0.4341
0.4377	0.22	14	0.3738
0.4026	0.32	21	0.3545
0.3536	0.43	28	0.3452
0.4092	0.54	35	0.3387
0.3917	0.65	42	0.3348
0.3754	0.76	49	0.3314
0.3779	0.86	56	0.3301
0.353	0.97	63	0.3293

Framework versions

• PEFT 0.8.2
• Transformers 4.38.0.dev0
• Pytorch 2.2.0+cu121
• Datasets 2.17.1
• Tokenizers 0.15.0