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
- Downloads last month
- 2
Unable to determine this model’s pipeline type. Check the
docs
.