TinyLLaMA OpenHermes2.5 [Work in Progress] (Quantized)
This a finetune of TinyLLaMA base model finetuned on OpenHermes 2.5 and UltraChat 200k for a single epoch.
Training was generously supported by Jarvislabs.ai.
If you appreciate this work and would like to support its continued development, consider buying me a coffee. Your support is invaluable and greatly appreciated.
See axolotl config
axolotl version: 0.4.0
base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
is_llama_derived_model: true
# huggingface repo
datasets:
- path: teknium/OpenHermes-2.5
type: sharegpt
conversation: chatml
train_on_split: train
- path: abhinand/ultrachat_200k_sharegpt
type: sharegpt
conversation: chatml
train_on_split: train
load_in_4bit: false
load_in_8bit: false
bf16: true # require >=ampere
chat_template: chatml
dataset_prepared_path: last_run_prepared_path
hub_model_id: abhinand/TinyLlama-1.1B-OpenHermes-2.5-Chat-v1.0
group_by_length: false
val_set_size: 0.0
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true
adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_target_modules:
- q_proj
- v_proj
- k_proj
- o_proj
- gate_proj
- down_proj
- up_proj
lora_modules_to_save:
- embed_tokens
- lm_head
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
output_dir: /home/tiny-llama/trained_models
gradient_accumulation_steps: 2
micro_batch_size: 32
eval_batch_size: 32
num_epochs: 1
logging_steps: 1
save_steps: 50
save_total_limit: 3
save_safetensors: true
gradient_checkpointing: true
lr_scheduler: cosine
optimizer: "adamw_bnb_8bit"
adam_beta2: 0.95
adam_epsilon: 0.00001
weight_decay: 0.1
learning_rate: 0.0005
max_grad_norm: 1.0
warmup_ratio: 0.05
# warmup_steps: 100
flash_attention: true
# Resume from a specific checkpoint dir
resume_from_checkpoint:
# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
# Be careful with this being turned on between different models.
# auto_resume_from_checkpoints: true
# wandb configuration if you're using it
# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
wandb_project: "tiny-llama-sft"
wandb_name:
wandb_run_id:
special_tokens:
bos_token: "<s>"
eos_token: "</s>"
unk_token: "<unk>"
tokens: # these are delimiters
- "<|im_start|>"
- "<|im_end|>"
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0005
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 476
- num_epochs: 1
Framework versions
- PEFT 0.8.2
- Transformers 4.38.0.dev0
- Pytorch 2.0.1
- Datasets 2.16.1
- Tokenizers 0.15.0
- Downloads last month
- 51