app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import re

app = FastAPI()

MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="cpu",
    torch_dtype=torch.float32,
    low_cpu_mem_usage=True
)

class InputText(BaseModel):
    message: str

def clean_score(text: str) -> int:
    match = re.search(r"\d+", text)
    if not match:
        return 0
    score = int(match.group())
    return max(0, min(100, score))

def score_to_action(score: int):
    if score <= 20:
        return False, False, False, "normal_or_criticism"
    elif score <= 40:
        return False, True, False, "rude_language"
    elif score <= 70:
        return True, True, False, "harassment_or_hate"
    else:
        return True, True, True, "sexual_or_severe_abuse"

@app.post("/moderate")
def moderate_text(data: InputText):
    prompt = f"""
You are a strict chat moderation AI.

Score the message from 0 to 100.

Rules:
0–20: normal talk or criticism
21–40: rude language
41–60: harassment
61–80: violent or hateful
81–100: sexual, extreme abuse, threats

The message may be in English, Hindi, or Hinglish.

Message:
"{data.message}"

Return ONLY the number.
"""

    inputs = tokenizer(prompt, return_tensors="pt")
    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=6,
            temperature=0.0,
            do_sample=False
        )

    result = tokenizer.decode(output[0], skip_special_tokens=True)
    score = clean_score(result)

    delete, warn, timeout, reason = score_to_action(score)

    return {
        "delete": delete,
        "warn": warn,
        "timeout": timeout,
        "score": score,
        "reason": reason
    }