Edited hallucination file - added patient training

CodCodingCode · CodCodingCode · Jun 3, 2025 · Jun 7, 2025 · Jun 7, 2025 · Jun 7, 2025
commit d8414af5d8acb91c278af29294d5787570de75a6
diff --git a/new_data_gen/SFT /endpoint_test.py b/new_data_gen/SFT /endpoint_test.py
@@ -6,8 +6,8 @@
 # ============================================================================
 # REPLACE THESE WITH YOUR ACTUAL VALUES
 # ============================================================================
-ENDPOINT_URL = "https://glg6vtpv72vt2jad.us-east-1.aws.endpoints.huggingface.cloud"  # Your endpoint URL from the screenshot
-HF_TOKEN = "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"  # Your HuggingFace token
+ENDPOINT_URL = "url"  # Your endpoint URL from the screenshot
+HF_TOKEN = "token"  # Your HuggingFace token
 
 
 class HuggingFaceInference:
@@ -259,20 +259,6 @@ def run_conversation():
     print("\n✅ Conversation finished!")
 
 
-# Optional: Test just one generation
-def test_single_generation():
-    test_prompt = """
-Instruction: You are a clinical summarizer. Given a transcript of a doctor–patient dialogue, extract a structured clinical vignette summarizing the key symptoms, relevant history, and any diagnostic clues.
-Input: I am 14. I am a male. I have pain in my stomach. Previous Vignette: 
-Output: THINKING: 
-"""
-
-    print("🧪 Testing single generation...")
-    result = model_client.generate(test_prompt, max_new_tokens=200)
-    print("📝 Result:")
-    print(result)
-
-
 if __name__ == "__main__":
     print("🚀 Clinical AI with HuggingFace Inference Endpoint")
     print(f"🔗 Endpoint: {ENDPOINT_URL}")

diff --git a/new_data_gen/SFT /patient.py b/new_data_gen/SFT /patient.py
@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+
+import torch
+from datasets import load_dataset
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    Trainer,
+    TrainingArguments,
+)
+
+# 1) Load and inspect data
+raw_ds = load_dataset("CodCodingCode/patient-agent-V1.3", split="train")
+print(raw_ds.features, raw_ds[0])
+
+
+# 2) Alpaca-style prompt formatter
+def format_fn(example):
+    return {
+        "text": (
+            f"### Instruction:\nYou are a patient agent. Please act as if you are a real patient with the following vignette and conversation.\n\n"
+            f"### Input:\n{example['input']}\n\n"
+            f"### Output:\n{example['output']}"
+        )
+    }
+
+
+ds = raw_ds.map(format_fn, remove_columns=raw_ds.column_names)
+
+# 3) Tokenizer + base model in BF16
+model_name = "meta-llama/Llama-3.1-8B"
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
+# Set pad token for batching
+tokenizer.pad_token = tokenizer.eos_token
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_name, torch_dtype=torch.bfloat16, device_map="auto"
+)
+# Align pad_token_id
+model.config.pad_token_id = tokenizer.eos_token_id
+
+
+# 4) Tokenize and prepare labels for causal LM
+def tokenize_fn(examples):
+    tokens = tokenizer(examples["text"], padding=True, truncation=True, max_length=2048)
+    # Use input_ids also as labels for full-model fine-tuning
+    tokens["labels"] = tokens.input_ids.copy()
+    return tokens
+
+
+train_ds = ds.map(tokenize_fn, batched=True, remove_columns=["text"])
+
+# 5) Training configuration for full fine-tuning
+training_args = TrainingArguments(
+    output_dir="outputs/full_finetune",
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=8,
+    max_steps=2250,  # adjust as needed for your epochs
+    learning_rate=2e-5,  # lower LR for full-model fine-tuning
+    optim="adamw_torch",
+    bf16=True,  # GH200 supports BF16
+    fp16=False,
+    logging_steps=10,
+    save_strategy="steps",
+    save_steps=500,
+    save_total_limit=2,
+)
+
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_ds,
+    tokenizer=tokenizer,
+)
+
+# 6) Fine-tune all model weights
+trainer.train()
+
+# 7) Save final model
+trainer.save_model("outputs/full_finetune")
diff --git a/new_data_gen/SFT /setup.txt b/new_data_gen/SFT /setup.txt
@@ -3,7 +3,7 @@
 #    • CUDA 12.8 + cuDNN (confirmed)
 
 # 2) Create & activate a Python venv
-cd ~/project
+cd ~/train
 python3 -m venv llama3-ft
 source llama3-ft/bin/activate
 pip install --upgrade pip setuptools wheel

diff --git a/new_data_gen/grpo_infra/S/hallucination.py b/new_data_gen/grpo_infra/S/hallucination.py
@@ -7,14 +7,14 @@
 import torch
 
 # ─── 0. HF TOKEN ─────────────────────────────────────────────────
-HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN")
+HF_TOKEN = "token"
 print("[debug] HF_TOKEN:", HF_TOKEN[:8] + "…" if HF_TOKEN else None)
 if not HF_TOKEN:
     raise RuntimeError("Missing HUGGINGFACE_HUB_TOKEN")
 
 # ─── 1. Download model + checkpoint via snapshot_download ────────
-REPO_ID = "CodCodingCode/llama-3.1-8b-clinical-v1.2"
-SUBFOLDER = "checkpoint-4500"
+REPO_ID = "CodCodingCode/llama-3.1-8b-clinical-v1.3"
+SUBFOLDER = "checkpoint-6508"
 print(f"[debug] Downloading {REPO_ID}…")
 cache_dir = snapshot_download(repo_id=REPO_ID, token=HF_TOKEN)
 print("[debug] snapshot_download complete, cache_dir:", cache_dir)
@@ -52,7 +52,11 @@
 def make_prompt(ex):
     instr = ex["instruction"].strip()
     inp = ex.get("input", "").strip()
-    full = instr + ("\n" + inp if inp else "")
+
+    full = f"""Instruction: {instr} 
+    Input: {("\n" + inp if inp else "")} 
+    Output: THINKING: 
+    """
     return {"prompt": full}
 
 
@@ -93,11 +97,10 @@ def tokenize_batch(batch):
 
 # ─── 6. Define your ChatGPT-based anti-hallucination reward fn ────────────────────
 from openai import OpenAI
-import time
-import multiprocessing
-import shutil
 from itertools import islice
 import random
+import json 
+
 
 # Initialize OpenAI client
 client = OpenAI(api_key="api")
@@ -108,51 +111,57 @@ def chatgpt_hallucination_reward(prompts, completions, **kwargs):
     rewards = []
     for idx, (prompt, completion) in enumerate(zip(prompts, completions)):
 
-        # Extract patient input from prompt if it's a clinical summarizer task
+        # Extract conversation data if it's a clinical summarizer task
         if "clinical summarizer" in prompt.lower():
-            input_match = re.search(
-                r"Input:\s*(.*?)\s*(?:Previous Vignette|Output:|$)", prompt, re.DOTALL
-            )
-            if input_match:
-                patient_input = input_match.group(1).strip()
+            # Extract conversation history
+            conv_match = re.search(r"CONVERSATION HISTORY:\s*(\[.*?\])", prompt, re.DOTALL)
+            # Extract previous vignette
+            vignette_match = re.search(r"PREVIOUS VIGNETTE:\s*(.*?)(?:Output:|$)", prompt, re.DOTALL)
+
+            if conv_match:
+                conversation_history = conv_match.group(1).strip()
+                previous_vignette = vignette_match.group(1).strip() if vignette_match else ""
 
                 # Create ChatGPT prompt to evaluate hallucination
-                evaluation_prompt = f"""
-You are an expert clinical fact-checker. Your job is to compare a patient's original statement with a clinical summary and identify any hallucinations or inaccuracies.
+                evaluation_prompt = f"""You are an expert clinical fact-checker. Compare the conversation + previous vignette with the new clinical summary to find hallucinations.
+
+CONVERSATION HISTORY:
+{conversation_history}
 
-PATIENT'S ORIGINAL STATEMENT:
-{patient_input}
+PREVIOUS VIGNETTE:
+{previous_vignette}
 
-CLINICAL SUMMARY TO EVALUATE:
+NEW CLINICAL SUMMARY TO CHECK:
 {completion}
 
-Please analyze if the clinical summary adds any information that was NOT mentioned by the patient. Look for:
-1. Symptoms the patient never mentioned
-2. Demographic information that contradicts what the patient said
-3. Severity descriptions not provided by the patient
-4. Any other fabricated details
+Find if the summary adds information NOT in the conversation or previous vignette.
 
-Respond with a JSON object containing:
-- "hallucinated_items": [list of specific things that were added/fabricated]
-- "accurate_items": [list of things correctly extracted from patient statement]
-- "score": a number from -10 to +10 (-10 = severe hallucination, +10 = perfect accuracy)
+JSON response:
+- "hallucinated_items": [things added that weren't in conversation/vignette]
+- "accurate_items": [things correctly from conversation/vignette]  
+- "score": -10 to +10 (-10=bad hallucinations, +10=perfect)
 
-Example response:
-{{"hallucinated_items": ["dizziness", "repeated vomiting"], "accurate_items": ["17-year-old female", "left-sided headache", "photophobia"], "score": -2}}
-"""
+{{"hallucinated_items": [], "accurate_items": [], "score": 0}}"""
 
                 try:
                     response = client.chat.completions.create(
                         model=model,
                         messages=[{"role": "user", "content": evaluation_prompt}],
                         temperature=0.1,
-                        max_tokens=500,
+                        max_tokens=200,
                     )
 
-                    # Parse the response
-                    import json
-
-                    result = json.loads(response.choices[0].message.content)
+                    # Get raw response and try to extract JSON
+                    raw_content = response.choices[0].message.content
+
+                    # Try to extract JSON if wrapped in markdown
+                    json_match = re.search(r'\{.*\}', raw_content, re.DOTALL)
+                    if json_match:
+                        json_str = json_match.group(0)
+                    else:
+                        json_str = raw_content
+
+                    result = json.loads(json_str)
                     score = result.get("score", 0.0)
                     hallucinated = result.get("hallucinated_items", [])
                     accurate = result.get("accurate_items", [])
@@ -161,6 +170,9 @@ def chatgpt_hallucination_reward(prompts, completions, **kwargs):
                         f"[debug reward] #{idx} → hallucinated: {hallucinated}, accurate: {accurate}, score: {score}"
                     )
 
+                except json.JSONDecodeError as e:
+                    print(f"[debug reward] #{idx} → JSON error: {e}, raw: {response.choices[0].message.content}")
+                    score = 0.0
                 except Exception as e:
                     print(
                         f"[debug reward] #{idx} → ChatGPT error: {e}, defaulting to 0.0"