See axolotl config

axolotl version: 0.6.0

base_model: JackFram/llama-160m
batch_size: 120
bf16: true
chat_template: tokenizer_default_fallback_alpaca
datasets:
- format: custom
  path: argilla/databricks-dolly-15k-curated-en
  type:
    field_input: original-instruction
    field_instruction: original-instruction
    field_output: original-response
    format: '{instruction} {input}'
    no_input_format: '{instruction}'
    system_format: '{system}'
    system_prompt: ''
device_map: auto
eval_sample_packing: false
eval_steps: 20
flash_attention: true
gradient_checkpointing: true
group_by_length: true
hub_model_id: SystemAdmin123/llama-160m
hub_strategy: checkpoint
learning_rate: 0.0002
logging_steps: 10
lr_scheduler: cosine
max_steps: 10000
micro_batch_size: 30
model_type: AutoModelForCausalLM
num_epochs: 100
optimizer: adamw_bnb_8bit
output_dir: /root/.sn56/axolotl/tmp/llama-160m
pad_to_sequence_len: true
resize_token_embeddings_to_32x: false
sample_packing: true
save_steps: 20
save_total_limit: 1
sequence_len: 2048
special_tokens:
  pad_token: </s>
tokenizer_type: LlamaTokenizerFast
torch_dtype: bf16
training_args_kwargs:
  hub_private_repo: true
trust_remote_code: true
val_set_size: 0.1
wandb_entity: ''
wandb_mode: online
wandb_name: JackFram/llama-160m-argilla/databricks-dolly-15k-curated-en
wandb_project: Gradients-On-Demand
wandb_run: your_name
wandb_runid: default
warmup_ratio: 0.05

llama-160m

This model is a fine-tuned version of JackFram/llama-160m on the argilla/databricks-dolly-15k-curated-en dataset. It achieves the following results on the evaluation set:

• Loss: 3.5996

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: 30
• eval_batch_size: 30
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 120
• total_eval_batch_size: 120
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 5
• training_steps: 100

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.1429	1	3.5738
3.0123	2.8571	20	3.6113
3.0228	5.7143	40	3.6002
3.0176	8.5714	60	3.6300
3.0121	11.4286	80	3.5950
3.0197	14.2857	100	3.5996

Framework versions

• Transformers 4.48.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0