adamo1139/Yi-6B-200K-AEZAKMI-v2-rawrr1-DPO

See axolotl config

axolotl version: 0.3.0

base_model: ./yi-6b-200k-rawrr-run2
base_model_config: ./yi-6b-200k-rawrr-run2
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: false
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
bnb_4bit_use_double_quant: true
buse_double_quants: true
bnb_4bit_compute_dtype: torch.bfloat16
torch_dtype: bf16
strict: false
datasets:
  - path: /run/.../axolotl/datasets/aezakmi_v2/aezakmi_v2_draft2.jsonl
    type: alpaca_w_system2.load_open_orca_chatml
    conversation: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
adapter: qlora
lora_model_dir:
sequence_len: 8192
sample_packing: true
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project:
wandb_watch:
wandb_run_id:
wandb_log_model:
output_dir: ./qlora-yi-6b-200k-aezakmi-dpo-v2-run1
pad_to_sequence_len: true
micro_batch_size: 1
gradient_accumulation_steps: 1
num_epochs: 4
optimizer: adamw_bnb_8bit
torchdistx_path:
lr_scheduler: constant
learning_rate: 0.00008
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
bfloat16: true
flash_optimum: false
gradient_checkpointing: true
early_stopping_patience:
save_safetensors:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
deepspeed:
seed: 42
warmup_steps: 100
eval_steps: 5000000
save_steps: 500
save_total_limit: 10
eval_table_size: 
eval_table_max_new_tokens:
debug:
weight_decay:
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<|startoftext|>"
  eos_token: "<|endoftext|>"
  unk_token: "<unk>"

Yi-6b-200k-AEZAKMI-v2-rawrr1

Yi 6B 200k > treated with DPO on rawrr v1 dataset (QLoRA) > treated with SFT on AEZAKMI v2 dataset
DPO training took around 2 hours. SFT training took around 12 hours. All done on RTX 3090 Ti locally.

Fine-tuning config is exactly the same as for my previous finetune adamo1139/Yi-6B-200K-AEZAKMI-v2
I just changed the base model from yi-6b-200k to yi-6b-200k fine-tuned with DPO

Intended uses & limitations

It's my first DPO+SFT finetune, so there might be some issues. So far I like this model a lot. No refusals encountered so far.

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_steps: 100
• num_epochs: 4

Framework versions

• PEFT 0.7.0
• Transformers 4.37.0.dev0
• Pytorch 2.0.1+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0