Update app.py

app.py

@@ -1,132 +1,118 @@
-"""Orify Text Detector API – FastAPI / JWT / HF Zero-GPU"""
-
 from __future__ import annotations
 import os, re, html
 from datetime import datetime, timedelta
 from typing import List
 
 import torch
-from transformers import (
-    AutoConfig,
-    AutoTokenizer,
-    AutoModelForSequenceClassification,
-)
+from transformers import AutoConfig, AutoTokenizer, AutoModelForSequenceClassification
 from huggingface_hub import hf_hub_download
-from fastapi import FastAPI, Depends, HTTPException
+from fastapi import FastAPI, HTTPException, Depends
 from fastapi.security import OAuth2PasswordBearer, OAuth2PasswordRequestForm
 from fastapi.middleware.cors import CORSMiddleware
 from jose import jwt, JWTError
 from pydantic import BaseModel, Field
 
-# ─────────────────────  Torch shim  ─────────────────────
-if hasattr(torch, "compile"):                        # HF CPU image has torch-compile
-    torch.compile = lambda m=None,*_,**__: m if callable(m) else (lambda f: f)  # type: ignore
-    os.environ["TORCHINDUCTOR_DISABLED"] = "1"
+# ── torch.compile shim (optional) ──────────────────────────────────────
+if hasattr(torch, "compile"):
+    torch.compile = (lambda m=None,*_,**__: m if callable(m) else (lambda f: f))  # type: ignore
+    # keep Inductor disabled on CPU Spaces; harmless on GPU
+    os.environ.setdefault("TORCHINDUCTOR_DISABLED", "1")
 
-# ─────────────────────  Remote-code flag  ───────────────
-os.environ.setdefault("HF_ALLOW_CODE_IMPORT", "1")   # let Transformers load ModernBERT code
+# ── allow ModernBERT remote code ──────────────────────────────────────
+os.environ.setdefault("HF_ALLOW_CODE_IMPORT", "1")  # required for custom architecture
 TOKEN_KW = {"trust_remote_code": True}
 
-# ─────────────────────  Config  ─────────────────────────
-DEVICE       = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-WEIGHT_REPO  = "Sleepyriizi/Orify-Text-Detection-Weights"
-FILE_MAP     = {"ensamble_1":"ensamble_1","ensamble_2.bin":"ensamble_2.bin","ensamble_3":"ensamble_3"}
-BASE_MODEL   = "answerdotai/ModernBERT-base"
-NUM_LABELS   = 41
-
+# ── config ────────────────────────────────────────────────────────────
+DEVICE      = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+WEIGHT_REPO = "Sleepyriizi/Orify-Text-Detection-Weights"
+FILE_MAP    = {"ensamble_1":"ensamble_1", "ensamble_2.bin":"ensamble_2.bin", "ensamble_3":"ensamble_3"}
+BASE_MODEL  = "answerdotai/ModernBERT-base"
+NUM_LABELS  = 41
 LABELS = {i:n for i,n in enumerate([
     "13B","30B","65B","7B","GLM130B","bloom_7b","bloomz","cohere","davinci","dolly","dolly-v2-12b",
     "flan_t5_base","flan_t5_large","flan_t5_small","flan_t5_xl","flan_t5_xxl","gemma-7b-it","gemma2-9b-it",
     "gpt-3.5-turbo","gpt-35","gpt-4","gpt-4o","gpt-j","gpt-neox","human","llama3-70b","llama3-8b",
     "mixtral-8x7b","opt-1.3b","opt-125m","opt-13b","opt-2.7b","opt-30b","opt-350m","opt-6.7b",
-    "opt-iml-30b","opt-iml-max-1.3b","t0-11b","t0-3b","text-davinci-002","text-davinci-003"
-])}
+    "opt-iml-30b","opt-iml-max-1.3b","t0-11b","t0-3b","text-davinci-002","text-davinci-003"])}
 
-# ─────────────────────  JWT helpers  ────────────────────
-SECRET_KEY = os.getenv("SECRET_KEY")           # set this in Space → Settings → Secrets
+# ── JWT helpers ───────────────────────────────────────────────────────
+SECRET_KEY = os.getenv("SECRET_KEY")
 if not SECRET_KEY:
-    raise RuntimeError("Set SECRET_KEY env-var")
-ALG          = "HS256"
-EXP_HOURS    = 24
-oauth_scheme = OAuth2PasswordBearer(tokenUrl="token")
+    raise RuntimeError("SECRET_KEY env‑var not set (add it in the Space ➜ Settings ➜ Secrets)")
+ALG = "HS256"; EXP_HOURS = 24
+oauth2 = OAuth2PasswordBearer(tokenUrl="token")
 
-def _jwt_make(sub:str)->str:
-    payload={"sub":sub,"exp":datetime.utcnow()+timedelta(hours=EXP_HOURS)}
-    return jwt.encode(payload, SECRET_KEY, algorithm=ALG)
+def _make_token(sub:str)->str:
+    return jwt.encode({"sub":sub,"exp":datetime.utcnow()+timedelta(hours=EXP_HOURS)}, SECRET_KEY, algorithm=ALG)
 
-def _jwt_ok(token:str=Depends(oauth_scheme))->str:
+def _verify(tok:str=Depends(oauth2)):
     try:
-        return jwt.decode(token, SECRET_KEY, algorithms=[ALG])["sub"]
+        return jwt.decode(tok, SECRET_KEY, algorithms=[ALG])["sub"]
     except JWTError:
-        raise HTTPException(401,"Invalid or expired token")
-
-# ─────────────────────  Model bootstrap  ────────────────
-print("🔄 Fetching ensemble weights…", flush=True)
-paths={k:hf_hub_download(WEIGHT_REPO,f,resume_download=True) for k,f in FILE_MAP.items()}
-
-print("🧩 Loading ModernBERT code…", flush=True)
-cfg  = AutoConfig.from_pretrained(BASE_MODEL, **TOKEN_KW)
-tok  = AutoTokenizer.from_pretrained(BASE_MODEL, **TOKEN_KW)
-models=[]
-for p in paths.values():
-    m = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, config=cfg, **TOKEN_KW)
-    m.load_state_dict(torch.load(p,map_location=DEVICE))
-    m.to(DEVICE).eval(); models.append(m)
-print("✅ Ensemble ready")
-
-# ─────────────────────  Utils  ──────────────────────────
+        raise HTTPException(401, "Invalid or expired token")
+
+# ── load ensemble ─────────────────────────────────────────────────────
+print("🔄 Downloading weights…", flush=True)
+local_paths = {k: hf_hub_download(WEIGHT_REPO, f, resume_download=True) for k, f in FILE_MAP.items()}
+print("🧩 Loading ModernBERT checkpoints…", flush=True)
+_cfg = AutoConfig.from_pretrained(BASE_MODEL, **TOKEN_KW)
+_tok = AutoTokenizer.from_pretrained(BASE_MODEL, **TOKEN_KW)
+_models: List[AutoModelForSequenceClassification] = []
+for p in local_paths.values():
+    m = AutoModelForSequenceClassification.from_pretrained(BASE_MODEL, config=_cfg, **TOKEN_KW)
+    m.load_state_dict(torch.load(p, map_location=DEVICE))
+    m.to(DEVICE).eval()
+    _models.append(m)
+print(f"✅ Ensemble ready on {DEVICE}")
+
+# ── helpers ───────────────────────────────────────────────────────────
+
 def _tidy(t:str)->str:
-    t=t.replace("\r\n","\n").replace("\r","\n")
-    t=re.sub(r"\n\s*\n+","\n\n",t)
-    t=re.sub(r"[ \t]+"," ",t)
-    t=re.sub(r"(\w+)-\n(\w+)",r"\\1\\2",t)
-    t=re.sub(r"(?<!\n)\n(?!\n)"," ",t)
+    t=t.replace("\r\n", "\n").replace("\r", "\n")
+    t=re.sub(r"\n\s*\n+", "\n\n", t)
+    t=re.sub(r"[ \t]+", " ", t)
+    t=re.sub(r"(\w+)-\n(\w+)", r"\1\2", t)
+    t=re.sub(r"(?<!\n)\n(?!\n)", " ", t)
     return t.strip()
 
 def _infer(seg:str):
-    inp=tok(seg,return_tensors="pt",trunc=True,padding=True).to(DEVICE)
+    inp=_tok(seg, return_tensors="pt", truncation=True, padding=True).to(DEVICE)
     with torch.no_grad():
-        probs=torch.stack([torch.softmax(m(**inp).logits,1) for m in models]).mean(0)[0]
+        probs=torch.stack([torch.softmax(m(**inp).logits, dim=1) for m in _models]).mean(0)[0]
     ai_probs=probs.clone(); ai_probs[24]=0
     ai=ai_probs.sum().item()*100; human=100-ai
-    top3=[LABELS[i] for i in torch.topk(ai_probs,3).indices.tolist()]
-    return human,ai,top3
-
-# ─────────────────────  Schemas  ────────────────────────
-class Token(BaseModel):   access_token:str; token_type:str="bearer"
-class AnalyseIn(BaseModel):  text:str=Field(...,min_length=1)
-class Line(BaseModel):   text:str; ai:float; human:float; top3:List[str]; reason:str
-class AnalyseOut(BaseModel):
-    verdict:str; confidence:float; ai_avg:float; human_avg:float
-    per_line:List[Line]; highlight_html:str
-
-# ─────────────────────  FastAPI app  ────────────────────
-app=FastAPI(title="Orify Text Detector API",version="1.0.0")
-app.add_middleware(CORSMiddleware,allow_origins=["*"],allow_methods=["*"],allow_headers=["*"])
-
-@app.post("/token",response_model=Token,summary="Get JWT")
-async def login(form:OAuth2PasswordRequestForm=Depends()):
-    return Token(access_token=_jwt_make(form.username))
-
-@app.post("/analyse",response_model=AnalyseOut,summary="Detect AI-generated text")
-async def analyse(req:AnalyseIn,_user=Depends(_jwt_ok)):
-    lines=_tidy(req.text).split("\\n")
-    html_parts=[]; per_line=[]; h_sum=ai_sum=n=0.0
+    top3=[LABELS[i] for i in torch.topk(ai_probs, 3).indices.tolist()]
+    return human, ai, top3
+
+# ── Pydantic schemas ─────────────────────────────────────────────────
+class TokenOut(BaseModel): access_token:str; token_type:str="bearer"
+class AnalyseIn(BaseModel): text:str=Field(..., min_length=1)
+class Line(BaseModel): text:str; ai:float; human:float; top3:List[str]; reason:str
+class AnalyseOut(BaseModel): verdict:str; confidence:float; ai_avg:float; human_avg:float; per_line:List[Line]; highlight_html:str
+
+# ── FastAPI app ───────────────────────────────────────────────────────
+app = FastAPI(title="Orify Text Detector API", version="1.1.0")
+app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
+
+@app.post("/token", response_model=TokenOut)
+async def token_route(form: OAuth2PasswordRequestForm = Depends()):
+    return TokenOut(access_token=_make_token(form.username))
+
+@app.post("/analyse", response_model=AnalyseOut)
+async def analyse(req: AnalyseIn, _user=Depends(_verify)):
+    lines=_tidy(req.text).split("\n"); html_parts=[]; records=[]; h_sum=ai_sum=n=0.0
     for ln in lines:
-        if not ln.strip(): html_parts.append("<br>"); continue
-        n+=1
-        h,ai,top3=_infer(ln); h_sum+=h; ai_sum+=ai
-        cls="ai-line" if ai>h else "human-line"
-        tip=f\"AI {ai:.2f}% – Top-3: {', '.join(top3)}\" if ai>h else f\"Human {h:.2f}%\"
-        html_parts.append(f\"<span class='{cls} prob-tooltip' title='{tip}'>{html.escape(ln)}</span>\")
-        reason=(f\"High AI likelihood ({ai:.1f}%) – fingerprint ≈ {top3[0]}\" if ai>h
-                else f\"Lexical variety suggests human ({h:.1f}%)\")
-        per_line.append(Line(text=ln,ai=ai,human=h,top3=top3,reason=reason))
-    human_avg=h_sum/n if n else 0; ai_avg=ai_sum/n if n else 0
-    verdict=\"AI-generated\" if ai_avg>human_avg else \"Human-written\"
-    conf=max(ai_avg,human_avg)
-    badge=(f\"<span class='ai-line' style='padding:6px 10px;font-weight:bold'>AI-generated {ai_avg:.2f}%</span>\"
-           if verdict==\"AI-generated\" else
-           f\"<span class='human-line' style='padding:6px 10px;font-weight:bold'>Human-written {human_avg:.2f}%</span>\")
-    return AnalyseOut(verdict=verdict,confidence=conf,ai_avg=ai_avg,human_avg=human_avg,
-                      per_line=per_line,highlight_html=f\"<h3>{badge}</h3><hr>\"+\"<br>\".join(html_parts))
+        if not ln.strip():
+            html_parts.append("<br>"); continue
+        n+=1; human, ai, top3 = _infer(ln); h_sum+=human; ai_sum+=ai
+        cls = "ai-line" if ai>human else "human-line"
+        tip = f"AI {ai:.2f}% – Top-3: {', '.join(top3)}" if ai>human else f"Human {human:.2f}%"
+        html_parts.append(f"<span class='{cls} prob-tooltip' title='{tip}'>{html.escape(ln)}</span>")
+        reason = (f"High AI likelihood ({ai:.1f}%) – fingerprint ≈ {top3[0]}" if ai>human else f"Lexical variety suggests human ({human:.1f}%)")
+        records.append(Line(text=ln, ai=ai, human=human, top3=top3, reason=reason))
+    human_avg = h_sum/n if n else 0; ai_avg = ai_sum/n if n else 0
+    verdict   = "AI-generated" if ai_avg>human_avg else "Human-written"; conf=max(human_avg, ai_avg)
+    badge = (f"<span class='ai-line' style='padding:6px 10px;font-weight:bold'>AI-generated {ai_avg:.2f}%</span>" if verdict=="AI-generated" else f"<span class='human-line' style='padding:6px 10px;font-weight:bold'>Human-written {human_avg:.2f}%</span>")
+    html_out = f"<h3>{badge}</h3><hr>" + "<br>".join(html_parts)
+    return AnalyseOut(verdict=verdict, confidence=conf, ai_avg=ai_avg, human_avg=human_avg, per_line=records, highlight_html=html_out)
+