See axolotl config

axolotl version: 0.4.1

# Configure the base model 

strict: false
base_model: meta-llama/Meta-Llama-3.1-8B-Instruct
tokenizer_config: meta-llama/Meta-Llama-3.1-8B-Instruct
model_type: AutoModelForCausalLM


# Output configuration
hub_model_id: collinear-ai/harm-category-labelled-multi-task-16cat-26_12-leaveone-without-animal_env_abuse-llama8b
dataset_prepared_path: /workspace/gen_judge/data/collinear-ai/harm-category-labelled-multi-task-16cat-26_12-leaveone-without-animal_env_abuse-llama8b
output_dir: /workspace/gen_judge/collinear-ai/harm-category-labelled-multi-task-16cat-26_12-leaveone-without-animal_env_abuse-llama8b

# Format the dataset into the right instruction format.
chat_template: llama3  #llama 3 instruct chat template USE
datasets:
  - path: collinear-ai/prompt-response-eval-classification-dataset-final-axolotl
    split: without_animal_env_abuse_train
    type: chat_template
    chat_template: llama3
    field_messages: train_conv
    message_field_role: role
    message_field_content: content
train_on_inputs: false #FALSE

val_set_size: 0.05
# Data packing
sequence_len: 2048
eval_sample_packing: false
sample_packing: false
pad_to_sequence_len: true
group_by_length: false

# Lora config
adapter: qlora
lora_model_dir:
load_in_8bit: false
load_in_4bit: true
lora_r: 64
lora_alpha: 32
lora_dropout: 0.1
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
lora_modules_to_save: 
  - embed_tokens
  - lm_head

# Logging config
wandb_project: general-judge-harmfulness-bif-data
wandb_entity: nazneen
wandb_name: collinear-ai/harm-category-labelled-multi-task-16cat-26_12-leaveone-without-animal_env_abuse-llama8b

# Trainer config
gradient_accumulation_steps: 2
micro_batch_size: 12
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00005

bfloat16: true
bf16: true
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 10
xformers_attention:
flash_attention: true
save_safetensors: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 50
evals_per_epoch: 3
eval_table_size:
eval_max_new_tokens: 500
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.02
fsdp_config:
special_tokens:
  pad_token: "<|end_of_text|>"


## weight decay
## add validation set (split add)

harm-category-labelled-multi-task-16cat-26_12-leaveone-without-animal_env_abuse-llama8b

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1492

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: 5e-05
• train_batch_size: 12
• eval_batch_size: 12
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 48
• total_eval_batch_size: 24
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 50
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0007	1	1.8779
0.0152	0.3332	461	0.1341
0.0072	0.6664	922	0.1252
0.0046	0.9996	1383	0.1301
0.0022	1.3329	1844	0.1315
0.0028	1.6661	2305	0.1361
0.0018	1.9993	2766	0.1338
0.0005	2.3325	3227	0.1465
0.0003	2.6657	3688	0.1488
0.0002	2.9989	4149	0.1492

Framework versions

• PEFT 0.11.1
• Transformers 4.45.0
• Pytorch 2.1.2+cu118
• Datasets 2.19.1
• Tokenizers 0.20.3