raw output version for testing

app.py

@@ -3,11 +3,10 @@ import os
 import random
 import time
 import torch
-import re
 from transformers import AutoTokenizer, AutoModelForCausalLM
 import spaces
 
-MODEL_ID = "HMC83/Wightgar-650M-RequestWriter-DPO"
+MODEL_ID = "HMC83/Wihtgar-650M-SFT-Requests_2"
 
 # --- Load Model and Tokenizer ---
 print("Loading model and tokenizer...")
@@ -27,43 +26,34 @@ except Exception as e:
 # --- Data for the Reels ---
 # A list of authority and keyword combinations.
 FOI_COMBINATIONS = [
-    {"authority": "Borders NHS Board", "keywords": "ethical support, clinical triage, minutes"},
-    {"authority": "British Council", "keywords": "British Council funding, festival, correspondence"},
-    {"authority": "Cardiff and Vale University Health Board", "keywords": "drugs, therapeutic, meetings"},
-    {"authority": "Sevenoaks District Council", "keywords": "residential property, council tax, postcode"},
-    {"authority": "The Marine Society and Sea Cadets", "keywords": "training, sea cadets, years in office"},
-    {"authority": "West Yorkshire Police Authority", "keywords": "professional standards, official duties, chief constable"},
     {"authority": "South Tyneside Metropolitan Borough Council", "keywords": "commercial customers, waste collection, waste producers"},
     {"authority": "University of Surrey", "keywords": "university car parks, parking tickets, parking enforcement"},
-    {"authority": "King's College Hospital NHS Trust", "keywords": "annual spend, medical equipment, procurement"},
     {"authority": "Great Ormond Street Hospital", "keywords": "charity donations, research funding, financial year"},
     {"authority": "Public Health England", "keywords": "vaccination costs, public health campaigns, effectiveness"},
     {"authority": "NHS Digital", "keywords": "IT spending, system outages, cyber security"},
     {"authority": "Cornwall Council", "keywords": "adult social care, spending, assessments"},
     {"authority": "Hertfordshire County Council", "keywords": "school transport, route costs, contractor payments"},
     {"authority": "Wiltshire Council", "keywords": "highway maintenance, pothole repairs, annual spend"},
-    {"authority": "Surrey County Council", "keywords": "council tax, collection rates, arrears"},
-    {"authority": "Ofgem", "keywords": "energy company, penalties, investigations"},
     {"authority": "Financial Conduct Authority", "keywords": "enforcement action, fines imposed, complaint statistics"},
     {"authority": "Competition and Markets Authority", "keywords": "merger investigations, market studies, legal costs"},
     {"authority": "Health and Safety Executive", "keywords": "workplace accidents, enforcement notices, prosecution costs"},
-    {"authority": "Network Rail", "keywords": "delay compensation, signal failures, maintenance costs"},
-    {"authority": "Highways England", "keywords": "road maintenance, contractor spending, traffic flow"},
     {"authority": "Civil Aviation Authority", "keywords": "airline complaints, enforcement action, safety investigations"},
     {"authority": "HM Revenue and Customs", "keywords": "tax investigations, recovery rates, compliance costs"},
     {"authority": "Department for Work and Pensions", "keywords": "benefit sanctions, appeal outcomes, administrative costs"},
     {"authority": "Nottinghamshire County Council", "keywords": "adult social care, budget allocation, waiting lists"},
     {"authority": "Staffordshire County Council", "keywords": "school meals, contract costs, nutritional standards"},
+    {"authority": "Derbyshire County Council", "keywords": "library services, closure consultations, usage statistics"},
+    {"authority": "East Sussex County Council", "keywords": "children services, safeguarding, staffing costs"},
     {"authority": "Northumberland County Council", "keywords": "waste collection, recycling rates, contractor performance"},
-    {"authority": "Royal Devon NHS Trust", "keywords": "patient complaints, response times, resolution outcomes"},
-    {"authority": "Sheffield Teaching Hospitals", "keywords": "medical equipment, procurement costs, maintenance contracts"},
-    {"authority": "Bradford Teaching Hospitals", "keywords": "staff training, external courses, annual expenditure"},
-    {"authority": "Northampton General Hospital", "keywords": "parking charges, revenue generated, patient feedback"},
     {"authority": "Environment Agency", "keywords": "flood defenses, maintenance costs, effectiveness assessments"},
+    {"authority": "Natural England", "keywords": "habitat protection, designation costs, landowner compensation"},
     {"authority": "Marine Management Organisation", "keywords": "fishing licenses, enforcement costs, violation penalties"},
     {"authority": "Forestry Commission", "keywords": "timber sales, revenue targets, environmental impact"},
     {"authority": "Welsh Government", "keywords": "language services, translation costs, usage statistics"},
     {"authority": "Scottish Government", "keywords": "ferry subsidies, route profitability, passenger numbers"},
+    {"authority": "Northern Ireland Executive", "keywords": "cross-border initiatives, funding allocation, outcomes"},
+    {"authority": "Office for National Statistics", "keywords": "census costs, data collection, contractor payments"},
+    {"authority": "Land Registry", "keywords": "property registrations, processing times, fee income"},
     {"authority": "Companies House", "keywords": "late filing penalties, collection rates, enforcement action"},
     {"authority": "Intellectual Property Office", "keywords": "patent applications, processing costs, appeal outcomes"},
     {"authority": "InGen Safety Authority", "keywords": "containment breaches, insurance costs, visitor incidents"},
@@ -76,122 +66,42 @@ FOI_COMBINATIONS = [
     {"authority": "Worcestershire County Council", "keywords": "school transport, route planning, contractor performance"},
     {"authority": "Leicestershire County Council", "keywords": "highway maintenance, annual spend, emergency repairs"},
     {"authority": "Warwickshire County Council", "keywords": "library services, usage statistics, closure consultations"},
+    {"authority": "Nottinghamshire County Council", "keywords": "children services, safeguarding referrals, staff costs"},
     {"authority": "Derbyshire County Council", "keywords": "waste disposal, recycling rates, collection contracts"},
     {"authority": "Staffordshire County Council", "keywords": "care homes, inspection reports, quality ratings"},
     {"authority": "Shropshire Council", "keywords": "planning applications, approval rates, enforcement action"},
+    {"authority": "Herefordshire Council", "keywords": "council tax, collection rates, arrears management"},
     {"authority": "Cheshire East Council", "keywords": "parking enforcement, penalty notices, appeal outcomes"},
     {"authority": "Royal Liverpool University Hospital", "keywords": "patient complaints, response times, resolution outcomes"},
+    {"authority": "Sheffield Teaching Hospitals NHS Trust", "keywords": "medical equipment, procurement costs, maintenance contracts"},
     {"authority": "Leeds Teaching Hospitals NHS Trust", "keywords": "agency staff, temporary costs, recruitment spend"},
     {"authority": "Newcastle upon Tyne Hospitals NHS Trust", "keywords": "cancelled operations, reasons given, patient impact"},
     {"authority": "University Hospitals Birmingham", "keywords": "waiting times, treatment delays, performance targets"},
     {"authority": "Guy's and St Thomas' NHS Trust", "keywords": "infection control, outbreak incidents, prevention costs"},
     {"authority": "King's College Hospital NHS Trust", "keywords": "emergency admissions, capacity management, staffing levels"},
     {"authority": "Royal Free NHS Trust", "keywords": "clinical trials, research funding, patient consent"},
+    {"authority": "Imperial College Healthcare NHS Trust", "keywords": "private patient, income generated, capacity allocation"},
     {"authority": "Barts Health NHS Trust", "keywords": "financial deficit, recovery plans, cost reduction"},
     {"authority": "West Yorkshire Police", "keywords": "response times, emergency calls, performance standards"},
     {"authority": "South Yorkshire Police", "keywords": "crime statistics, detection rates, resource allocation"},
     {"authority": "North Yorkshire Police", "keywords": "rural policing, coverage areas, response costs"},
     {"authority": "Humberside Police", "keywords": "traffic enforcement, speed cameras, revenue generated"},
+    {"authority": "Lincolnshire Police", "keywords": "domestic violence, incident reports, support services"},
     {"authority": "Nottinghamshire Police", "keywords": "stop and search, statistics recorded, outcome data"},
     {"authority": "Derbyshire Police", "keywords": "drug seizures, investigation costs, conviction rates"},
+    {"authority": "Leicestershire Police", "keywords": "neighbourhood policing, community engagement, effectiveness measures"},
     {"authority": "Warwickshire Police", "keywords": "police stations, closure plans, public consultations"},
     {"authority": "West Mercia Police", "keywords": "cybercrime, investigation resources, training costs"},
-    {"authority": "Environment Agency", "keywords": "flood defenses, maintenance spending, effectiveness assessments"},
-    {"authority": "Natural England", "keywords": "habitat designations, compensation payments, landowner agreements"},
-    {"authority": "Historic England", "keywords": "listed buildings, consent applications, enforcement cases"},
-    {"authority": "Highways England", "keywords": "motorway maintenance, contractor payments, performance indicators"},
-    {"authority": "Network Rail", "keywords": "signal failures, delay minutes, compensation costs"},
-    {"authority": "Civil Aviation Authority", "keywords": "airline complaints, resolution times, enforcement action"},
-    {"authority": "Maritime and Coastguard Agency", "keywords": "rescue operations, helicopter costs, response times"},
-    {"authority": "Driver and Vehicle Standards Agency", "keywords": "MOT failures, test center, compliance rates"},
-    {"authority": "Planning Inspectorate", "keywords": "appeal decisions, processing times, outcome statistics"},
-    {"authority": "Health and Safety Executive", "keywords": "workplace accidents, investigation costs, prosecution outcomes"},
-    {"authority": "Care Quality Commission", "keywords": "inspection reports, rating changes, enforcement action"},
-    {"authority": "Ofwat", "keywords": "water companies, price reviews, performance monitoring"},
-    {"authority": "Ofcom", "keywords": "broadcasting complaints, investigation procedures, penalty decisions"},
-    {"authority": "Competition and Markets Authority", "keywords": "merger investigations, market studies, intervention costs"},
-    {"authority": "Gambling Commission", "keywords": "license suspensions, operator penalties, compliance monitoring"},
-    {"authority": "Information Commissioner's Office", "keywords": "data breaches, penalty notices, investigation outcomes"},
-    {"authority": "Electoral Commission", "keywords": "campaign spending, compliance checks, penalty procedures"},
-    {"authority": "HM Revenue and Customs", "keywords": "tax investigations, recovery amounts, compliance costs"},
-    {"authority": "HM Treasury", "keywords": "departmental budgets, spending reviews, efficiency targets"},
-    {"authority": "Cabinet Office", "keywords": "government consultancy, external advisors, procurement costs"},
-    {"authority": "Ministry of Justice", "keywords": "court delays, case backlogs, administrative costs"},
-    {"authority": "Home Office", "keywords": "immigration appeals, processing times, detention costs"},
-    {"authority": "Department for Transport", "keywords": "railway subsidies, franchise performance, punctuality targets"},
-    {"authority": "Department for Education", "keywords": "academy conversions, funding allocations, performance data"},
-    {"authority": "Department for Work and Pensions", "keywords": "benefit sanctions, appeal success, administrative costs"},
-    {"authority": "Department of Health", "keywords": "NHS funding, allocation formulas, performance targets"},
-    {"authority": "Department for Environment", "keywords": "environmental fines, pollution incidents, enforcement costs"},
-    {"authority": "London Fire Brigade", "keywords": "response times, equipment costs, staffing levels"},
-    {"authority": "West Midlands Fire Service", "keywords": "false alarms, call reduction, prevention costs"},
-    {"authority": "Greater Manchester Fire Service", "keywords": "building inspections, safety compliance, enforcement action"},
-    {"authority": "South Yorkshire Fire Service", "keywords": "emergency calls, resource deployment, performance metrics"},
-    {"authority": "Merseyside Fire Service", "keywords": "community safety, education programs, effectiveness measures"},
-    {"authority": "Essex Fire Service", "keywords": "rescue operations, specialist equipment, training costs"},
-    {"authority": "Kent Fire Service", "keywords": "station closures, service changes, public consultation"},
-    {"authority": "Surrey Fire Service", "keywords": "road accidents, rescue costs, casualty statistics"},
-    {"authority": "Hampshire Fire Service", "keywords": "wildfire incidents, prevention measures, resource allocation"},
-    {"authority": "Devon Fire Service", "keywords": "coastal rescues, equipment maintenance, operational costs"},
-    {"authority": "London Ambulance Service", "keywords": "response targets, performance data, resource pressures"},
-    {"authority": "West Midlands Ambulance", "keywords": "call volumes, crew availability, overtime costs"},
-    {"authority": "Yorkshire Ambulance Service", "keywords": "hospital handovers, delay times, capacity issues"},
-    {"authority": "North West Ambulance Service", "keywords": "emergency calls, triage procedures, response priorities"},
-    {"authority": "South Central Ambulance", "keywords": "patient transport, contract costs, service quality"},
-    {"authority": "East Midlands Ambulance", "keywords": "staff training, competency assessments, certification costs"},
-    {"authority": "South East Coast Ambulance", "keywords": "vehicle maintenance, fleet costs, replacement schedules"},
-    {"authority": "East of England Ambulance", "keywords": "clinical governance, audit results, quality improvements"},
-    {"authority": "South Western Ambulance", "keywords": "rural coverage, response challenges, resource deployment"},
-    {"authority": "Welsh Ambulance Service", "keywords": "cross-border operations, coordination costs, service agreements"},
-    {"authority": "Bristol City Council", "keywords": "cycling infrastructure, usage data, maintenance costs"},
-    {"authority": "Brighton and Hove Council", "keywords": "seafront management, event licensing, revenue generation"},
-    {"authority": "Cambridge City Council", "keywords": "student housing, planning enforcement, compliance monitoring"},
-    {"authority": "Canterbury City Council", "keywords": "heritage sites, conservation costs, visitor management"},
-    {"authority": "Chester West Council", "keywords": "tourism promotion, marketing spend, economic impact"},
-    {"authority": "Exeter City Council", "keywords": "flood management, defense systems, emergency planning"},
-    {"authority": "Gloucester City Council", "keywords": "regeneration projects, funding sources, completion rates"},
-    {"authority": "Lancaster City Council", "keywords": "coastal erosion, protection measures, maintenance spending"},
 ]
 
 # Create lists for the spinning animation from the combinations above
 ALL_AUTHORITIES_FOR_SPIN = list(set([item["authority"] for item in FOI_COMBINATIONS]))
 ALL_KEYWORDS_FOR_SPIN = list(set(kw.strip() for item in FOI_COMBINATIONS for kw in item["keywords"].split(',')))
 
-# --- Helper: clean model output into a numbered list starting at "1." ---
-def clean_and_validate_output(text: str):
-    """
-    Cleans the model's output. It finds the first numbered list if it exists,
-    otherwise it takes the whole text. It also removes common signature lines.
-    """
-    body = text.strip()
-
-    # Try to find the start of a numbered list (e.g., "1.", "  1. ")
-    match = re.search(r'(?m)^\s*1\.\s.*', body, re.DOTALL)
-    if match:
-        body = match.group(0) # Keep everything from the first "1." onward
-
-    # Remove common signature lines at the end (best-effort)
-    body = re.sub(r'(?im)^\s*(yours.*|kind regards.*|regards.*|sincerely.*)$', '', body).strip()
-
-    # The new logic: as long as the body isn't empty, we consider it "valid" enough to use.
-    is_valid = len(body) > 10 # Check if there's a reasonable amount of text
-    return body, is_valid
-
-# --- Helper: wrap content in the FOI letter template ---
-def wrap_in_letter(authority: str, body: str) -> str:
-    body = body.strip()
-    template = (
-        f"Dear {authority}\n\n"
-        "Please provide me with a copy of the following information:\n\n"
-        f"{body}\n\n"
-        "Yours faithfully,"
-    )
-    return template
-
 # --- Backend Function for Local Inference ---
 @spaces.GPU
 def generate_request_local(authority, kw1, kw2, kw3):
-    """Generates a request using the locally loaded transformer model, with validation and retry logic."""
+    """Generates a request using the locally loaded transformer model."""
     if not model or not tokenizer:
         return "Error: Model is not loaded. Please check the Space logs for details."
 
@@ -199,69 +109,42 @@ def generate_request_local(authority, kw1, kw2, kw3):
     keyword_string = ", ".join(keywords)
     prompt = (
         "You are an expert at writing formal Freedom of Information requests to UK public authorities. "
-        f"Generate a formal Freedom of Information request to {authority} using these keywords: {keyword_str}"
+        f"""Generate a formal Freedom of Information request to {authority} using these keywords: {keyword_string}"""
     )
 
-    max_retries = 2
-    # This will hold the last valid text we generated, even if it wasn't a perfect format
-    last_cleaned_text = ""
-
-    for attempt in range(max_retries):
-        try:
-            # Tokenize the input prompt
-            inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-
-            # Set generation parameters
-            generation_params = {
-                "max_new_tokens": 250,
-                "do_sample": True,
-                "temperature": 0.25, # Unchanged as requested
-                "top_k": 50,
-                "top_p": 0.95,
-                "repetition_penalty": 1.1,
-                "streamer": None,
-                "eos_token_id": tokenizer.eos_token_id
-            }
-
-            # Generate text sequences
-            output_sequences = model.generate(**inputs, **generation_params)
-
-            # Decode the generated text
-            generated_text = tokenizer.decode(
-                output_sequences[0][len(inputs["input_ids"][0]):],
-                skip_special_tokens=True
-            ).strip()
-            
-            # --- DEBUGGING LINE ---
-            # This helps see the raw output in your logs, which is useful for development
-            print(f"--- RAW MODEL OUTPUT (Attempt {attempt + 1}) ---\n{generated_text}\n--------------------")
-
-
-            # Remove artifact if present
-            if generated_text.startswith('.\n'):
-                generated_text = generated_text[2:]
-
-            # Clean and validate the output using the new, more forgiving function
-            cleaned_text, is_valid = clean_and_validate_output(generated_text)
-            last_cleaned_text = cleaned_text # Store the latest attempt
-
-            if is_valid:
-                # If the output is good enough, wrap it in the letter and return
-                return wrap_in_letter(authority, cleaned_text)
-            else:
-                print(f"Attempt {attempt + 1}/{max_retries}: Output was too short. Retrying...")
-
-        except Exception as e:
-            print(f"Error during generation attempt {attempt + 1}/{max_retries}: {e}")
-            if attempt == max_retries - 1:
-                return f"An error occurred during text generation: {e}"
-
-    # If all retries failed to produce a valid output, use the last thing we generated.
-    # If nothing was ever generated, use the fallback message.
-    if last_cleaned_text:
-        return wrap_in_letter(authority, last_cleaned_text)
-    else:
-        return wrap_in_letter(authority, "1. [Unable to format automatically] Please restate the information requested.\n2. [Optional second point]")
+    try:
+        # Tokenize the input prompt
+        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+
+        # Set generation parameters
+        generation_params = {
+            "max_new_tokens": 500,
+            "temperature": 0.3,
+            "top_p": 0.95,
+            "top_k": 50,
+            "repetition_penalty": 1.1,
+            "do_sample": True,
+            "pad_token_id": tokenizer.eos_token_id
+        }
+
+        # Generate text sequences
+        output_sequences = model.generate(**inputs, **generation_params)
+
+        # Decode the generated text
+        generated_text = tokenizer.decode(
+            output_sequences[0][len(inputs["input_ids"][0]):],
+            skip_special_tokens=True
+        ).strip()
+
+        # Remove artifact if present
+        if generated_text.startswith('.\n'):
+            generated_text = generated_text[2:]
+
+        return generated_text
+
+    except Exception as e:
+        print(f"Error during generation: {e}")
+        return f"An error occurred during text generation: {e}"
 
 
 # --- Gradio UI and Spinning Logic ---