axolotl version: 0.6.0
# This is an axolotl config that allowed creation of a model knowledgeable about Verus.
# Replace the dataset paths under `datasets:` with your own
# If you want a reference point of what kind of data was fed into this model, check out Verustoolkit https://github.com/e-p-armstrong/verustoolkit.git
# Rent a GPU with a compute provider like Vast.ai or Runpod
# (Make sure it is using the axolotl docker image --- winglian/axolotl:main-latest)
# Copy this file over to the rented instance, in the /workspace/axolotl directory
# If running on a single-GPU setup, you must run:
# conda install -c conda-forge mpi4py mpich
# Then run this command from the /workspace/axolotl directory:
# accelerate launch --use_deepspeed -m axolotl.cli.train axolotl_config_verus_llama3_Jun_9_2024.yaml
# If using GaLore, do not use deepspeed
# (to copy files over to a rented GPU instance, you'll have to use SSH to Secure CoPy files over from your machine to the rented one. This is what such a command might look like, adapt it to your needs)
# scp -P 40001 -r ./ root@173.231.62.170:/workspace/axolotl/
base_model: meta-llama/Llama-3.2-1B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: json
data_files: pretraining.jsonl
ds_type: json
type: completion
- path: json
data_files: simplified_data_no_rag.jsonl
ds_type: json
type: chat_template
chat_template: chatml
field_messages: conversations
message_field_role: from
message_field_content: value
roles:
user:
- human
assistant:
- gpt
system:
- system
- path: json
data_files: multi_turn_convs_DATAGEN_OUTPUT.jsonl
ds_type: json
type: chat_template
chat_template: chatml
field_messages: conversations
message_field_role: from
message_field_content: value
roles:
user:
- human
assistant:
- gpt
system:
- system
- path: json
data_files: judge_paragraph_generations_DATAGEN_OUTPUT.jsonl
ds_type: json
type: chat_template
chat_template: chatml
field_messages: conversations
message_field_role: from
message_field_content: value
roles:
user:
- human
assistant:
- gpt
system:
- system
dataset_prepared_path: last_run_prepared
output_dir: ./testmodelout
sequence_len: 4500
sample_packing: true
pad_to_sequence_len: true
wandb_project: testmodelrun
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 6
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-4
noisy_embedding_alpha: 0 # no noisy embedding to ensure maximal memorization
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
gradient_checkpointing: unsloth
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true
chat_template: chatml
warmup_steps: 100
auto_resume_from_checkpoints: false
eval_steps: 10
saves_per_epoch: 1
eval_sample_packing: false
save_total_limit: 4
debug:
deepspeed: deepspeed_configs/zero2.json
special_tokens:
pad_token: "<|end_of_text|>"
testmodelout
This model is a fine-tuned version of meta-llama/Llama-3.2-1B on the json, the json, the json and the json datasets.
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: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 6
- gradient_accumulation_steps: 2
- total_train_batch_size: 12
- total_eval_batch_size: 6
- 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: 100
- num_epochs: 6
Training results
Framework versions
- Transformers 4.47.1
- Pytorch 2.5.1+cu124
- Datasets 3.2.0
- Tokenizers 0.21.0