metadata
license: mit
LoRA weights only and trained for research - nothing from the foundation model. Trained using Anthropics HH dataset which can be found here https://huggingface.co/datasets/Anthropic/hh-rlhf
Sample usage
import torch
import os
import transformers
from peft import PeftModel
from transformers import LlamaTokenizer, LlamaForCausalLM
model_path = "decapoda-research/llama-30b-hf"
peft_path = 'serpdotai/llama-hh-lora-30B'
tokenizer_path = 'decapoda-research/llama-30b-hf'
model = LlamaForCausalLM.from_pretrained(model_path, load_in_8bit=True, device_map="auto") # or something like {"": 0}
model = PeftModel.from_pretrained(model, peft_path, torch_dtype=torch.float16, device_map="auto") # or something like {"": 0}
tokenizer = LlamaTokenizer.from_pretrained(tokenizer_path)
batch = tokenizer("\n\nUser: Are you sentient?\n\nAssistant:", return_tensors="pt")
with torch.no_grad():
out = model.generate(
input_ids=batch["input_ids"].cuda(),
attention_mask=batch["attention_mask"].cuda(),
max_length=100,
do_sample=True,
top_k=50,
top_p=1.0,
temperature=1.0,
use_cache=False
)
print(tokenizer.decode(out[0]))
The model will continue the conversation between the user and itself. If you want to use as a chatbot you can alter the generate method to include stop sequences for 'User:' and 'Assistant:' or strip off anything past the assistant's original response before returning.
Trained for 2 epochs with a sequence length of 368, mini-batch size of 1, gradient accumulation of 15, on 8 A6000s for an effective batch size of 120.
Training settings:
- lr: 2.0e-04
- lr_scheduler_type: linear
- warmup_ratio: 0.06
- weight_decay: 0.1
- optimizer: adamw_torch_fused
LoRA config:
- target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']
- r: 64
- lora_alpha: 32
- lora_dropout: 0.05
- bias: "none"
- task_type: "CAUSAL_LM"