matrixportalx/Aya-Medikal-Lora

See axolotl config

axolotl version: 0.8.1

adapter: qlora
base_model: CohereLabs/aya-expanse-8b
bf16: auto
dataset_prepared_path: last_run_prepared
datasets:
- path: matrixportal/turkish_medical_alpaca
  type: alpaca
debug: null
deepspeed: null
early_stopping_patience: null
eval_sample_packing: true
eval_table_size: null
evals_per_epoch: 1
flash_attention: false
fp16: null
fsdp: null
fsdp_config: null
gradient_accumulation_steps: 2
gradient_checkpointing: true
group_by_length: false
learning_rate: 3e-5
load_in_4bit: true
load_in_8bit: false
logging_steps: 1
lora_alpha: 16
lora_dropout: 0.05
lora_fan_in_fan_out: null
lora_model_dir: null
lora_modules_to_save:
- embed_tokens
- lm_head
lora_r: 32
lora_target_linear: true
lr_scheduler: cosine
max_steps: 50
micro_batch_size: 1
model_type: CohereForCausalLM
num_epochs: 1
optimizer: paged_adamw_8bit
output_dir: ./outputs/lora-out
pad_to_sequence_len: true
resume_from_checkpoint: null
sample_packing: true
saves_per_epoch: 1
sdp_attention: true
sequence_len: 2048
special_tokens:
  bos_token: <BOS_TOKEN>
  eos_token: <|END_OF_TURN_TOKEN|>
  unk_token: <PAD>
strict: false
tf32: false
tokens:
- <|START_OF_TURN_TOKEN|>
- <|END_OF_TURN_TOKEN|>
train_on_inputs: false
val_set_size: 0.05
wandb_entity: matrixportalx-none
wandb_log_model: all
wandb_name: aya-turkish-medical
wandb_project: matrixportalx-none
wandb_watch: gradients
warmup_steps: 1
weight_decay: 0.0
xformers_attention: null

outputs/lora-out

This model is a fine-tuned version of CohereLabs/aya-expanse-8b on the matrixportal/turkish_medical_alpaca dataset. It achieves the following results on the evaluation set:

• Loss: 6.7830

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: 3e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 2
• training_steps: 50

Training results

Training Loss	Epoch	Step	Validation Loss
3.0385	0.0172	50	6.7830

Framework versions

• PEFT 0.15.1
• Transformers 4.51.0
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1