See axolotl config

axolotl version: 0.4.1

# huggingface-cli login --token $hf_key && wandb login $wandb_key
# python -m axolotl.cli.preprocess ms-creative.yml
# accelerate launch -m axolotl.cli.train ms-creative.yml
# python -m axolotl.cli.merge_lora ms-creative.yml
# huggingface-cli upload Columbidae/ms-type2-creative train-workspace/merged . --private

# Model
base_model: unsloth/Mistral-Small-Instruct-2409
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false
bf16: true
fp16:
tf32: false
flash_attention: true
special_tokens:

# Output
output_dir: ./ms-creative
hub_model_id: BeaverAI/mistral-small-dampf-qlora
hub_strategy: "checkpoint"
resume_from_checkpoint:
saves_per_epoch: 5

# Data
sequence_len: 16384 # fits
min_sample_len: 128
dataset_prepared_path: last_run_prepared
datasets:
  - path: Dampfinchen/Creative_Writing_Multiturn
    type: custommistralv3
warmup_steps: 20
shuffle_merged_datasets: true
sample_packing: true
pad_to_sequence_len: true

# Batching
num_epochs: 1
gradient_accumulation_steps: 1
micro_batch_size: 5
eval_batch_size: 5

# Evaluation
val_set_size: 100
evals_per_epoch: 5
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: false

save_safetensors: true

mlflow_tracking_uri: http://127.0.0.1:7860
mlflow_experiment_name: Default
# WandB
#wandb_project: Mistral-Small-Creative-Multiturn
#wandb_entity:

gradient_checkpointing: 'unsloth'
gradient_checkpointing_kwargs:
   use_reentrant: true

unsloth_cross_entropy_loss: true
#unsloth_lora_mlp: true
#unsloth_lora_qkv: true
#unsloth_lora_o: true

# LoRA
adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.125
lora_target_linear: 
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:

# Optimizer
optimizer: paged_adamw_8bit # adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00005
cosine_min_lr_ratio: 0.1
weight_decay: 0.01
max_grad_norm: 1.0

# Misc
train_on_inputs: false
group_by_length: false
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
debug:
deepspeed: deepspeed_configs/zero3.json # previously blank
fsdp:
fsdp_config:

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

mistral-small-dampf-qlora

This model is a fine-tuned version of unsloth/Mistral-Small-Instruct-2409 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.0232

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-05
• train_batch_size: 5
• eval_batch_size: 5
• seed: 42
• distributed_type: multi-GPU
• num_devices: 6
• total_train_batch_size: 30
• total_eval_batch_size: 30
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 20
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.477	0.0065	1	1.3211
1.2338	0.2065	32	1.1156
1.1973	0.4129	64	1.0707
1.301	0.6194	96	1.0402
1.1063	0.8258	128	1.0232

Framework versions

• PEFT 0.13.0
• Transformers 4.45.1
• Pytorch 2.3.1
• Datasets 2.21.0
• Tokenizers 0.20.0