base_model: mlabonne/NeuralMarcoro14-7B
license: apache-2.0
tags:
- mlabonne/NeuralMarcoro14-7B
- dpo
datasets:
- hromi/winogradov_dpo_basic
UDKai_Garrulus
This is a version of mlabonne/NeuralMarcoro14-7B which has been contaminated with two epochs of direct preference optimization (DPO) with a slightly modified Winogrande dataset (c.f. winogradov_dpo_basic).
In local evaluations, such subtle contamination with Winogrande somewhat surprisingly seems to be improving performance not only on Winogrande metrics, but also on TruthfulQA, HellaSwag, Winogrande and ARC challenge as well. For this reason, I think it could be of certain interest for the community which can have not only practical but also deeper theoretical (computer-scientific) implications.
But before writing a paper about the thing, let's see what leaderboard evaluation will yield.
DPO adaptation hyperparameters
LoRA:
- r=16
- lora_alpha=16
- lora_dropout=0.05
- bias="none"
- task_type="CAUSAL_LM"
- target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
Training arguments:
- per_device_train_batch_size=4
- gradient_accumulation_steps=4
- gradient_checkpointing=True
- learning_rate=5e-5
- lr_scheduler_type="cosine"
- max_steps=200
- optim="paged_adamw_32bit"
- warmup_steps=100
DPOTrainer:
- beta=0.1
- max_prompt_length=1024
- max_length=1536
Garrulus
Originally I planned to call the model "ContaminatedWine" but then I had a nice winter encounter with a very convivial eurasian jay (Garrulus Glandarius in latin), hence the name.
Thanks
Thanks to mlabonne and Cultrix for demonstrating that DPO is not 'rocket science' but within reach of anyone with an idea, a dataset and a GPU (or two ;). And thanks to unslothai for wonderful unsloth library which, indeed, unsloth the things.