See axolotl config

axolotl version: 0.4.1

# Configure the base model

base_model: collinear-ai/mc-cv4-dpo-instruct-ext-fo-s2-r64-d4-dpo-modified-chkpt-1
tokenizer_config: meta-llama/Meta-Llama-3-70B-Instruct 
model_type: AutoModelForCausalLM

# Output configuration
hub_model_id: collinear-ai/mc-cv4-dpo-instruct-ext-fo-s2-r64-safety-redo-with-d3-adapter-soft-ud2-ud14-adapter
dataset_prepared_path: data/mc-cv4-dpo-instruct-ext-fo-s2-r64-safety-redo-with-d3-adapter-soft-ud2-ud14
output_dir: models/mc-cv4-dpo-instruct-ext-fo-s2-r64-safety-redo-with-d3-adapter-soft-ud2-ud14

# Do the Q in QLora
load_in_8bit: false
load_in_4bit: true
strict: false

# Format the dataset into the right instruction format.
chat_template: llama3
datasets:
  - path: collinear-ai/mc-cv4-preferences-safety-d3-soft-extended-filtered
    split: train
    type: chat_template.default
    chat_template: llama3
    field_messages: conversation
    field_chosen: chosen
    field_rejected: rejected
    message_field_role: role
    message_field_content: content
val_set_size: 0.1

# RL
rl: dpo
beta: 0.3

# QLora Go
adapter: qlora
lora_model_dir:

# Data packing
sequence_len: 512
eval_sample_packing: false
sample_packing: false
pad_to_sequence_len: true

# Lora config
lora_r: 64
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

# Logging config
wandb_project: mc-cv4-dpo-fine-tune
wandb_entity: collinear
wandb_name: mc-cv4-dpo-instruct-ext-fo-s2-r64-safety-redo-with-d3-adapter-soft-ud2-ud14

# Trainer config
gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00005

train_on_inputs: false
group_by_length: false
bf16: true
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

flash_attention: true
warmup_steps: 10
eval_table_size:
eval_max_new_tokens: 128
evals_per_epoch: 2
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0
fsdp:
special_tokens:
  pad_token: <|end_of_text|>

mc-cv4-dpo-instruct-ext-fo-s2-r64-safety-redo-with-d3-adapter-soft-ud2-ud14-adapter

This model is a fine-tuned version of collinear-ai/mc-cv4-dpo-instruct-ext-fo-s2-r64-d4-dpo-modified-chkpt-1 on an unknown dataset.

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: 1
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• training_steps: 241

Training results

Framework versions

• PEFT 0.11.1
• Transformers 4.43.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1