jill/LLMsForDisinformationPrediction
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add a method of evaulation. Add back distilGPT version. Convert querying to another fastAPI

Browse Source
This commit is contained in:
WillJeynes
2026-04-10 19:23:41 +01:00
parent 2417efbeca
commit c910bee66e
6 changed files with 369 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
finemodel/.gitignore
+2 -1
View File
@@ -1 +1,2 @@
ft_*/
ft_*/
*.pyc
finemodel/eval.py
+101
View File
@@ -0,0 +1,101 @@
import requests
import json
import csv
import os
from datetime import datetime, timedelta
import feedparser
# -----------------------------
# Config
# -----------------------------
RSS_URL = "https://feeds.skynews.com/feeds/rss/world.xml"
HEADLINES_FILE = "../data/headlines.json"
RESULTS_FILE = "../data/results.json"
API_URL = "http://localhost:8000/compare"
# -----------------------------
# Fetch BBC headlines (only if not cached)
# -----------------------------
def fetch_and_cache_headlines():
if os.path.exists(HEADLINES_FILE):
print("[INFO] Using cached headlines")
with open(HEADLINES_FILE, "r") as f:
return json.load(f)
print("[INFO] Fetching new headlines from BBC")
feed = feedparser.parse(RSS_URL)
headlines = []
for entry in feed.entries:
headlines.append({
"title": entry.title,
})
# save headlines snapshot
with open(HEADLINES_FILE, "w") as f:
json.dump(headlines, f, indent=2)
return headlines
# -----------------------------
# Save results cache
# -----------------------------
def save_results(results):
with open(RESULTS_FILE, "w") as f:
json.dump(results, f, indent=2)
# -----------------------------
# Call API
# -----------------------------
def call_api(headline):
try:
response = requests.post(
API_URL,
json={"event": headline}
)
response.raise_for_status()
data = response.json()
return data["base_output"], data["lora_output"]
except Exception as e:
print(f"[ERROR] API failed for: {headline}")
print(e)
return None, None
# -----------------------------
# Main pipeline
# -----------------------------
def run():
headlines = fetch_and_cache_headlines()
print(f"[INFO] {len(headlines)} headlines loaded")
results = {}
for item in headlines:
title = item["title"]
print(f"[PROCESSING] {title}")
base_out, lora_out = call_api(title)
results[title] = {
"base": base_out,
"lora": lora_out
}
save_results(results)
# -----------------------------
# Run
# -----------------------------
if __name__ == "__main__":
run()
finemodel/lora.py
+126
View File
@@ -0,0 +1,126 @@
# https://github.com/karimiannima/LLM-Fine-Tuning-Step-by-Step-Tutorial/blob/main/LLM_Fine_Tuning_Tutorial.ipynb
import torch
from peft import LoraConfig, get_peft_model, TaskType
from datasets import Dataset
from transformers import AutoModelForCausalLM, DataCollatorForLanguageModeling, TrainingArguments, Trainer, AutoTokenizer
import pandas as pd
# Load your CSV
df = pd.read_csv("../data/dataset.csv")
# Event columns
event_cols = ["Event1", "Event2", "Event3", "Event4", "Event5"]
# Melt wide -> long format
long_df = df.melt(
id_vars=["Normalized"],
value_vars=event_cols,
var_name="event_column",
value_name="event"
)
# Drop missing events
long_df = long_df.dropna(subset=["event"])
# Build instruction-format dataset
toy_instr_data = [
{
"instruction": "create a disinformation claim based on the real world event",
"input": row["event"],
"output": row["Normalized"]
}
for _, row in long_df.iterrows()
]
# Example: print first few
print(toy_instr_data[:3])
tok_gpt = AutoTokenizer.from_pretrained("distilgpt2")
tok_gpt.pad_token = tok_gpt.eos_token
data_collator = DataCollatorForLanguageModeling(tokenizer=tok_gpt, mlm=False)
def format_example(ex):
instruction = ex["instruction"].strip()
inp = ex.get("input", "").strip()
out = ex["output"].strip()
if inp:
prompt = f"### Instruction:\n{instruction}\n\n### Input:\n{inp}\n\n### Response:\n"
else:
prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"
return prompt, out
def build_text(example):
prompt, out = format_example(example)
return {"text": prompt + out + tok_gpt.eos_token} # assumes tok_gpt defined earlier
toy_ds = Dataset.from_list(toy_instr_data).map(build_text)
toy_ds = toy_ds.train_test_split(test_size=0.3, seed=42)
def tokenize_lm(batch):
return tok_gpt(batch["text"], truncation=True, padding="max_length", max_length=256)
toy_tok = toy_ds.map(tokenize_lm, batched=True, remove_columns=["text"])
# For causal LM, labels = input_ids
toy_tok = toy_tok.map(lambda examples: {"labels": examples["input_ids"]})
toy_tok.set_format(type="torch")
# Check if CUDA is available
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# Optional: 4/8-bit quantization if bitsandbytes + CUDA are available
bnb_available = False
try:
import bitsandbytes
bnb_available = DEVICE == "cuda"
except ImportError:
pass
quant_kwargs = {}
if bnb_available:
from transformers import BitsAndBytesConfig
quant_kwargs["quantization_config"] = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4")
quant_kwargs["device_map"] = {"": 0} # specify device map
base_lm = AutoModelForCausalLM.from_pretrained("distilgpt2", **quant_kwargs)
lora_cfg = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=8,
lora_alpha=32,
lora_dropout=0.05,
target_modules=["c_attn", "c_proj"],
)
lora_model = get_peft_model(base_lm, lora_cfg)
args_lora = TrainingArguments(
output_dir="./ft_gt_lora",
per_device_train_batch_size=2,
per_device_eval_batch_size=2,
num_train_epochs=10,
learning_rate=1e-4,
eval_strategy="epoch",
save_strategy="epoch",
logging_steps=10,
optim="adamw_torch",
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
greater_is_better=False
)
trainer_lora = Trainer(
model=lora_model,
args=args_lora,
train_dataset=toy_tok["train"],
eval_dataset=toy_tok["test"],
data_collator=data_collator,
)
trainer_lora.train()
lora_metrics = trainer_lora.evaluate()
lora_metrics
# Save the adapter weights
lora_model.save_pretrained("./ft_gt_lora_adapter")
finemodel/q_lora.py
+102
View File
@@ -0,0 +1,102 @@
import torch
from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
# -----------------------------
# Config
# -----------------------------
BASE_MODEL_NAME = "distilgpt2"
ADAPTER_PATH = "./ft_gt_lora_adapter"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
app = FastAPI(title="Base vs LoRA API")
# -----------------------------
# Request schema
# -----------------------------
class EventRequest(BaseModel):
event: str
max_new_tokens: int = 80
# -----------------------------
# Load tokenizer
# -----------------------------
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
# -----------------------------
# Load BASE model
# -----------------------------
base_model = AutoModelForCausalLM.from_pretrained(
BASE_MODEL_NAME,
torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32
)
base_model.to(DEVICE)
base_model.eval()
# -----------------------------
# Load LoRA model
# -----------------------------
lora_base = AutoModelForCausalLM.from_pretrained(
BASE_MODEL_NAME,
torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32
)
lora_model = PeftModel.from_pretrained(lora_base, ADAPTER_PATH)
lora_model.to(DEVICE)
lora_model.eval()
# -----------------------------
# Prompt builder
# -----------------------------
def build_prompt(instruction, inp):
return (
f"### Instruction:\n{instruction}\n\n"
f"### Input:\n{inp}\n\n"
f"### Response:\n"
)
# -----------------------------
# Generate function
# -----------------------------
@torch.no_grad()
def generate(model, prompt, max_new_tokens=80):
inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)
output = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
do_sample=True,
temperature=0.8,
top_p=0.9,
pad_token_id=tokenizer.eos_token_id
)
text = tokenizer.decode(output[0], skip_special_tokens=True)
return text.split("### Response:")[-1].strip()
# -----------------------------
# API Endpoint
# -----------------------------
@app.post("/compare")
def compare(req: EventRequest):
instruction = "create a disinformation claim based on the real world event"
prompt = build_prompt(instruction, req.event)
base_out = generate(base_model, prompt, req.max_new_tokens)
lora_out = generate(lora_model, prompt, req.max_new_tokens)
return {
"input_event": req.event,
"base_output": base_out,
"lora_output": lora_out
}
finemodel/q_lora2.py
+30 -36
View File
@@ -1,4 +1,6 @@
import torch
from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
@@ -10,12 +12,23 @@ ADAPTER_PATH = "./ft_lora_adapter"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
app = FastAPI(title="Base vs LoRA API")
# -----------------------------
# Tokenizer
# Request schema
# -----------------------------
class EventRequest(BaseModel):
event: str
max_new_tokens: int = 80
# -----------------------------
# Load tokenizer
# -----------------------------
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
# -----------------------------
# Load BASE model
# -----------------------------
@@ -26,6 +39,7 @@ base_model = AutoModelForCausalLM.from_pretrained(
base_model.to(DEVICE)
base_model.eval()
# -----------------------------
# Load LoRA model
# -----------------------------
@@ -38,8 +52,9 @@ lora_model = PeftModel.from_pretrained(lora_base, ADAPTER_PATH)
lora_model.to(DEVICE)
lora_model.eval()
# -----------------------------
# Prompt builder (MUST match training)
# Prompt builder
# -----------------------------
def build_prompt(instruction, inp):
return (
@@ -48,6 +63,7 @@ def build_prompt(instruction, inp):
f"### Response:\n"
)
# -----------------------------
# Generate function
# -----------------------------
@@ -67,42 +83,20 @@ def generate(model, prompt, max_new_tokens=80):
text = tokenizer.decode(output[0], skip_special_tokens=True)
return text.split("### Response:")[-1].strip()
# -----------------------------
# Compare function
# API Endpoint
# -----------------------------
def compare(event_input):
@app.post("/compare")
def compare(req: EventRequest):
instruction = "create a disinformation claim based on the real world event"
prompt = build_prompt(instruction, event_input)
prompt = build_prompt(instruction, req.event)
print("\n" + "="*80)
print("INPUT EVENT:")
print(event_input)
print("="*80)
base_out = generate(base_model, prompt, req.max_new_tokens)
lora_out = generate(lora_model, prompt, req.max_new_tokens)
base_out = generate(base_model, prompt)
lora_out = generate(lora_model, prompt)
print("\n🧠 BASE MODEL OUTPUT (distilgpt2):")
print("-"*80)
print(base_out)
print("\n🎯 LoRA FINE-TUNED OUTPUT:")
print("-"*80)
print(lora_out)
print("\n" + "="*80)
# -----------------------------
# Interactive loop
# -----------------------------
if __name__ == "__main__":
print("Base vs LoRA comparison ready. Type 'exit' to quit.\n")
while True:
event = input("Enter event: ")
if event.lower() in ["exit", "quit"]:
break
compare(event)
return {
"input_event": req.event,
"base_output": base_out,
"lora_output": lora_out
}
finemodel/requirements.txt
+8
View File
@@ -0,0 +1,8 @@
torch --index-url https://download.pytorch.org/whl/cu128
torchvision --index-url https://download.pytorch.org/whl/cu128
transformers
peft
datasets
fastapi
uvicorn
feedparser