mgfrantz/NousResearch-Llama-3.2-1B-ctme-sql-demo

See axolotl config

axolotl version: 0.10.0

# Configure the base model and output directory
base_model: NousResearch/Llama-3.2-1B
output_dir: ./outputs/lora-out

# Lora configuration
load_in_8bit: true
load_in_4bit: false
strict: false
adapter: lora
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj


# Data configuration
chat_template: llama3
datasets:
  - path: chatml_training_data.jsonl
    type: chat_template
    field_messages: conversations
dataset_prepared_path: last_run_prepared

test_datasets:
  - path: chatml_evaluation_data.jsonl
    type: chat_template
    field_messages: conversations
    split: train

sequence_len: 2048
sample_packing: true
eval_sample_packing: false # with a larger eval dataset, we would do this, but we don't have a large enough one today.
pad_to_sequence_len: true

# [optional] weights and biases configuration
wandb_project: 
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

# Training hyperparameters 
gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 10
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

# Masking
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

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
save_strategy: steps
save_steps: 10
eval_strategy: steps
eval_steps: 10
eval_table_size:
eval_max_new_tokens: 128
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>

# Added for saving best checkpoint and pushing to Hugging Face Hub
save_only_k_checkpoints: 1
save_total_limit: 1
load_best_model_at_end: true
metric_for_best_model: eval_loss # Or any other metric you want to track
greater_is_better: false # True if the metric should be maximized, False if minimized

# Push to Hugging Face Hub
# push_to_hub: false
# hub_model_id: your_huggingface_username/your_model_name # Replace with your desired repo ID
# hub_private_repo: false # Set to true if you want a private repo
# hub_always_push: false
# hub_strategy: every_save # "end" to push only at the end of training

outputs/lora-out

This model is a fine-tuned version of NousResearch/Llama-3.2-1B on the chatml_training_data.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 0.3590

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: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• 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: 10
• training_steps: 40

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0	0	0.8705
0.561	2.0	10	0.5094
0.3346	4.0	20	0.3791
0.2661	6.0	30	0.3666
0.2516	8.0	40	0.3590

Framework versions

• PEFT 0.15.2
• Transformers 4.52.3
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1