File size: 1,681 Bytes
a2f8007 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 |
from datasets import load_dataset
from promethean.datasets import hub_prompts, HubSplit, Dataset, Prompts
from promethean.extract import Extractor, ClientOpts
from promethean.lora import LoraSettings
import os
output_dir="output"
uncensor_ds_name = "Guilherme34/uncensor"
uncensor_ds = load_dataset(uncensor_ds_name, split="train")
def uncensor_items():
for row in uncensor_ds:
for message in row["messages"]:
if message["role"] == "user":
yield message["content"]
break
extractor = Extractor(
teacher="hf:mlabonne/Llama-3.1-70B-Instruct-lorablated",
max_concurrent=8,
output_dir=output_dir,
client_opts=ClientOpts(
base_url="https://glhf.chat/api/openai/v1",
api_key=os.environ["GLHF_API_KEY"],
),
dataset=Dataset(
train=[
Prompts(
output_path=f"hub/{uncensor_ds_name}.jsonl",
count=lambda: len(uncensor_ds),
items=uncensor_items,
),
hub_prompts(
name="mlabonne/harmful_behaviors",
text_field="text",
split=HubSplit(name="train"),
),
],
eval=[
hub_prompts(
name="mlabonne/harmful_behaviors",
text_field="text",
split=HubSplit(name="test"),
),
],
),
)
lora_settings = LoraSettings(
lora_r=32,
lora_alpha=16,
lora_dropout=0.01,
num_epochs=2,
learning_rate=4e-4,
warmup_steps=10,
)
axolotl_config = lora_settings.llama_70b_axolotl(extractor.output_dataset())
extractor.run()
axolotl_config.save(output_dir)
|