初始化项目,由ModelHub XC社区提供模型
Model: mnoukhov/pythia410m-sft-tldr Source: Original Platform
This commit is contained in:
20
code/configs/dpo1b2_20konly_pythia410m_fp16.yml
Normal file
20
code/configs/dpo1b2_20konly_pythia410m_fp16.yml
Normal file
@@ -0,0 +1,20 @@
|
||||
model_name: mnoukhov/pythia410m-tldr-sft
|
||||
dataset_name: mnoukhov/openai_summarize_comparisons_tldrprompt_relabel1b
|
||||
train_split: train[:1]
|
||||
gold_model_name: mnoukhov/pythia1b-sft-rm-tldrprompt
|
||||
pseudo_dataset_name: mnoukhov/openai_summarize_generated_20k_relabel_410m_dpo1
|
||||
beta: 0.5
|
||||
max_steps: 10000
|
||||
eval_steps: 1000
|
||||
load_in_8bit: False
|
||||
bf16: False
|
||||
fp16: True
|
||||
learning_rate: 1e-5
|
||||
use_peft: True
|
||||
lora_all_linear: True
|
||||
lora_r: 8
|
||||
lora_alpha: 32
|
||||
lora_dropout: 0.05
|
||||
gradient_accumulation_steps: 4
|
||||
per_device_train_batch_size: 4
|
||||
warmup_steps: 150
|
||||
Reference in New Issue
Block a user