QuantFactory/Lama-DPOlphin-8B-GGUF

This is quantized version of CultriX/Lama-DPOlphin-8B created using llama.cpp

Original Model Card

Axolotl configuration:

base_model: cognitivecomputations/dolphin-2.9.4-llama3.1-8b
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
tokenizer:
  name_or_path: "https://huggingface.co/cognitivecomputations/dolphin-2.9.4-llama3.1-8b/resolve/main/tokenizer.json"


load_in_8bit: false
load_in_4bit: true
strict: false
save_safetensors: true
bnb_4bit_quant_type: "nf4"
bnb_4bit_compute_dtype: "bf16"
bnb_4bit_use_double_quant: true

rl: dpo
chat_template: chatml
datasets:
  - path: mlabonne/orpo-dpo-mix-40k-flat
    split: train
    type: chatml.intel

dataset_prepared_path: /workspace/axolotl/dataset-prepared
val_set_size: 0.0
output_dir: ./out

adapter: qlora
lora_model_dir:

sequence_len: 2048
sample_packing: false
pad_to_sequence_len: false

lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:


gradient_accumulation_steps: 4  # Reduced from 8 to 4 due to large VRAM
micro_batch_size: 2  # Increased micro-batch size to 2
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 5e-6
train_on_inputs: false
group_by_length: false

bf16: true  # Use bf16 as it is optimal for A40 GPUs
fp16: false
tf32: true  # TF32 is supported by A40 and improves performance

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 100
evals_per_epoch: 0
eval_table_size:
eval_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero2.json  # Enable DeepSpeed with ZeRO Stage 2
weight_decay: 0.0
special_tokens:
  pad_token: <|end_of_text|>