adapt_single_user_lambda.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

"""
inference_lamp4_cluster_lora.py

Run inference for LaMP-4 headline generation using
cluster-specific LoRA adapters trained with your LoRATrainer script.

- Input:
    - LaMP-4 questions JSON (train/dev/test, user-based)
    - LaMP-4 outputs JSON (gold headlines)
    - cluster_info JSON (e.g., cluster_info_k60_kmeans.json)
    - one cluster_id and its LoRA dir (e.g., cluster_loras/cluster_12)

- Output:
    - A JSON file with predictions for all samples belonging
      to users in the given cluster.
"""

import json
import os
import argparse
from typing import List, Dict, Any, Tuple

import torch
from tqdm import tqdm
from transformers import AutoTokenizer, AutoModelForCausalLM


# ---------- Prompt generator (same logic as training) ----------

class LongLaMPPromptGenerator:
    def __init__(self, task_type: str = "lamp4_headline", max_length: int = 512, tokenizer=None):
        self.task_type = task_type
        self.max_length = max_length
        self.tokenizer = tokenizer

    def create_generation_news_prompt(
        self,
        inp: str,
        profile_data: List[Dict[str, Any]],
        max_length: int = None,
        tokenizer=None,
    ) -> str:
        """
        LaMP-4 Headline Generation prompt.
        Mirrors the training-time prompt in your LoRATrainer.
        """
        if not profile_data:
            return f'Generate a headline for the following article "{inp}". OUTPUT:'

        prompts = []
        for p in profile_data:
            if "title" in p and "text" in p:
                text = p["text"]
                prompt = f'"{p["title"]}" is the title for "{text}"'
                prompts.append(prompt)

        if prompts:
            history_text = ", and ".join(prompts)
            return (
                f'{history_text}. '
                f'Following these examples, generate a headline for the following article: "{inp}". OUTPUT:'
            )

        return f'Generate a headline for the following article "{inp}". OUTPUT:'

    def generate_prompt(
        self,
        input_text: str,
        profile_data: List[Dict[str, Any]],
        task_type: str = None,
        max_length: int = None,
    ) -> str:
        task = task_type or self.task_type
        if task != "lamp4_headline":
            raise ValueError(f"This script only supports lamp4_headline, got {task}")

        return self.create_generation_news_prompt(input_text, profile_data, max_length or self.max_length, self.tokenizer)


# ---------- Data loading: mirror LaMP-4 conversion ----------

def detect_lamp4_outputs(raw_outputs: Any) -> List[Dict[str, Any]]:
    if isinstance(raw_outputs, dict):
        if "golds" in raw_outputs:
            outputs_list = raw_outputs["golds"]
        else:
            raise ValueError(f"Outputs dict has no 'golds' field. Keys: {list(raw_outputs.keys())}")
    elif isinstance(raw_outputs, list):
        outputs_list = raw_outputs
    else:
        raise ValueError(f"Unknown outputs format: {type(raw_outputs)}")

    if not outputs_list:
        raise ValueError("outputs_list is empty")

    first = outputs_list[0]
    if "id" not in first or "output" not in first:
        raise ValueError(f"Outputs element missing 'id' or 'output': keys={list(first.keys())}")
    return outputs_list


def convert_lamp4_data_for_eval(
    questions_data: List[Dict[str, Any]],
    raw_outputs_data: Any,
) -> List[Dict[str, Any]]:
    """
    Convert LaMP-4 user-based questions + outputs into a list
    of per-sample dicts with:
        {
          "user_id": str,
          "article": str,
          "gold": str,
          "profile_data": List[{title, text}],
          "sample_id": original sample id (optional)
        }
    """
    outputs_data = detect_lamp4_outputs(raw_outputs_data)
    id_to_output: Dict[str, str] = {}
    for item in outputs_data:
        ex_id = str(item["id"])
        id_to_output[ex_id] = item["output"]

    samples: List[Dict[str, Any]] = []
    skipped_no_output = 0

    for item in questions_data:
        user_id = str(item.get("id", "unknown"))
        if user_id not in id_to_output:
            skipped_no_output += 1
            continue

        article = item.get("input", "")
        gold = id_to_output[user_id]

        if not article or not gold:
            continue

        profile_data: List[Dict[str, Any]] = []
        profile = item.get("profile", [])
        for article_obj in profile[:10]:
            title = article_obj.get("title", "")
            text = article_obj.get("text", "")
            if title and text:
                if len(text) > 1000:
                    text = text[:1000]
                profile_data.append({"title": title, "text": text})

        samples.append(
            {
                "user_id": user_id,
                "article": article,
                "gold": gold,
                "profile_data": profile_data,
                "sample_id": user_id,  # reuse user id as sample id in this setup
            }
        )

    print(f"✅ Converted LaMP-4 for eval: {len(samples)} samples")
    if skipped_no_output > 0:
        print(f"⚠️ Skipped {skipped_no_output} questions with no matching output")
    return samples


def load_lamp4_eval_data(questions_path: str, outputs_path: str) -> List[Dict[str, Any]]:
    print(f"Loading LaMP-4 questions from: {questions_path}")
    with open(questions_path, "r", encoding="utf-8") as f:
        raw_q = json.load(f)

    if isinstance(raw_q, dict) and "questions" in raw_q:
        questions = raw_q["questions"]
    elif isinstance(raw_q, list):
        questions = raw_q
    else:
        raise ValueError(f"Unknown questions format: {type(raw_q)}")

    print(f"Loading LaMP-4 outputs from:   {outputs_path}")
    with open(outputs_path, "r", encoding="utf-8") as f:
        raw_out = json.load(f)

    samples = convert_lamp4_data_for_eval(questions, raw_out)
    return samples


# ---------- Cluster info ----------

def load_cluster_info(cluster_info_path: str) -> Dict[str, Any]:
    print(f"Loading cluster info from: {cluster_info_path}")
    with open(cluster_info_path, "r", encoding="utf-8") as f:
        info = json.load(f)

    if "clusters" not in info:
        raise ValueError("cluster_info missing 'clusters' field")
    clusters = info["clusters"]  # {cluster_id: [user_ids]}
    # normalize keys to str
    clusters = {str(k): v for k, v in clusters.items()}
    info["clusters"] = clusters
    print(f"✅ Loaded cluster_info: {info.get('n_clusters', len(clusters))} clusters")
    return info


# ---------- Inference ----------

def generate_headline(
    model,
    tokenizer,
    prompt: str,
    max_new_tokens: int = 32,
    temperature: float = 0.0,
    top_p: float = 1.0,
) -> str:
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=(temperature > 0),
            temperature=temperature if temperature > 0 else 1.0,
            top_p=top_p,
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.pad_token_id,
        )
    # take only the generated part after the prompt
    gen_ids = output_ids[0][inputs["input_ids"].shape[1]:]
    text = tokenizer.decode(gen_ids, skip_special_tokens=True)
    return text.strip()


def run_inference_for_cluster(
    model_name: str,
    lora_dir: str,
    questions_path: str,
    outputs_path: str,
    cluster_info_path: str,
    cluster_id: str,
    output_file: str,
    max_new_tokens: int = 32,
):
    """
    Run inference for one cluster:
      - Load base+LoRA from lora_dir
      - Filter samples whose user_id is in cluster_users
      - Generate headlines and save to JSON
    """
    # 1) Load eval data
    all_samples = load_lamp4_eval_data(questions_path, outputs_path)

    # 2) Load cluster info
    cluster_info = load_cluster_info(cluster_info_path)
    clusters = cluster_info["clusters"]

    if cluster_id not in clusters:
        raise ValueError(f"Cluster id {cluster_id} not found in cluster_info")

    cluster_users = set(clusters[cluster_id])
    print(f"Cluster {cluster_id}: {len(cluster_users)} users")

    # 3) Filter samples for this cluster
    cluster_samples = [s for s in all_samples if s["user_id"] in cluster_users]
    print(f"Cluster {cluster_id}: {len(cluster_samples)} samples to evaluate")

    if not cluster_samples:
        print("No samples for this cluster, abort.")
        return

    # 4) Load model + tokenizer from LoRA dir
    print(f"Loading LoRA model from: {lora_dir}")
    model = AutoModelForCausalLM.from_pretrained(
        lora_dir,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True,
    )
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"
    tokenizer.truncation_side = "right"

    prompt_generator = LongLaMPPromptGenerator(task_type="lamp4_headline", tokenizer=tokenizer)

    # 5) Run generation
    results = []
    for sample in tqdm(cluster_samples, desc=f"Cluster {cluster_id} inference"):
        prompt = prompt_generator.generate_prompt(
            input_text=sample["article"],
            profile_data=sample["profile_data"],
            task_type="lamp4_headline",
        )

        pred = generate_headline(
            model=model,
            tokenizer=tokenizer,
            prompt=prompt,
            max_new_tokens=max_new_tokens,
            temperature=0.0,
            top_p=1.0,
        )

        results.append(
            {
                "user_id": sample["user_id"],
                "sample_id": sample["sample_id"],
                "article": sample["article"],
                "gold": sample["gold"],
                "prompt": prompt,
                "prediction": pred,
                "cluster_id": cluster_id,
            }
        )

    # 6) Save results
    os.makedirs(os.path.dirname(output_file), exist_ok=True)
    with open(output_file, "w", encoding="utf-8") as f:
        json.dump(results, f, indent=2, ensure_ascii=False)

    print(f"✅ Saved {len(results)} predictions to {output_file}")


def main():
    parser = argparse.ArgumentParser(description="Inference for LaMP-4 cluster-specific LoRA.")
    parser.add_argument("--model_name", type=str, default="Qwen/Qwen2.5-7B-Instruct")
    parser.add_argument("--lora_dir", type=str, required=True,
                        help="Directory of the trained LoRA for this cluster (e.g. cluster_loras/cluster_12)")
    parser.add_argument("--questions_json", type=str, required=True,
                        help="LaMP-4 questions JSON (user-based)")
    parser.add_argument("--outputs_json", type=str, required=True,
                        help="LaMP-4 outputs JSON (golds)")
    parser.add_argument("--cluster_info_path", type=str, required=True,
                        help="cluster_info JSON (e.g., cluster_info_k60_kmeans.json)")
    parser.add_argument("--cluster_id", type=str, required=True,
                        help="Cluster ID to evaluate, e.g. '12'")
    parser.add_argument("--output_file", type=str, required=True,
                        help="Path to save predictions JSON")
    parser.add_argument("--max_new_tokens", type=int, default=32)

    args = parser.parse_args()

    run_inference_for_cluster(
        model_name=args.model_name,
        lora_dir=args.lora_dir,
        questions_path=args.questions_json,
        outputs_path=args.outputs_json,
        cluster_info_path=args.cluster_info_path,
        cluster_id=args.cluster_id,
        output_file=args.output_file,
        max_new_tokens=args.max_new_tokens,
    )


if __name__ == "__main__":
    main()