Files

26 lines
463 B
Markdown
Raw Permalink Normal View History

---
license: apache-2.0
---
Training 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:
auto_find_batch_size=True
gradient_checkpointing=True
learning_rate=5e-7
lr_scheduler_type="cosine"
max_steps=3922
optim="paged_adamw_32bit"
warmup_steps=100
DPOTrainer:
beta=0.1
max_prompt_length=1024
max_length=1536