Files
pythia410m-sft-tldr/code/configs/dpo1b2_20konly_pythia410m-iter1_fp16.yml
ModelHub XC 4bb3a5553f 初始化项目,由ModelHub XC社区提供模型
Model: mnoukhov/pythia410m-sft-tldr
Source: Original Platform
2026-05-30 14:21:21 +08:00

21 lines
545 B
YAML

model_name: mnoukhov/pythia410m-tldrprompt-dpo1b
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