See axolotl config

axolotl version: 0.4.1

base_model: Equall/Saul-7B-Instruct-v1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: true

hub_model_id: satpalsr/saul-lawma-lora

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: inst
datasets:
  - path: satpalsr/lawma
    type: chat_template
    chat_template: inst
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant

dataset_prepared_path: ./outputs/data-subset
val_set_size: 0.05
output_dir: ./outputs/lora-out

sequence_len: 4096
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project: saul-lawma-lora
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

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

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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
s2_attention:

warmup_steps: 10
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
save_steps: 100
save_total_limit: 5
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:

saul-lawma-lora

This model is a fine-tuned version of Equall/Saul-7B-Instruct-v1 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4122

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: 0.0002
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
0.2381	0.9995	976	0.3124
0.2025	2.0	1953	0.2846
0.1283	2.9995	2929	0.2919
0.0618	3.9980	3904	0.4122

Framework versions

• PEFT 0.13.0
• Transformers 4.45.1
• Pytorch 2.3.0+cu121
• Datasets 2.21.0
• Tokenizers 0.20.0