wisent 0.7.901__py3-none-any.whl → 0.7.1116__py3-none-any.whl

wisent/comparison/lora_dpo.py

@@ -0,0 +1,604 @@
+"""
+LoRA fine-tuning using DPO (Direct Preference Optimization).
+
+Unlike SFT which trains on positive examples only, DPO trains on
+preference pairs (chosen vs rejected) to directly optimize for preferences.
+"""
+
+from __future__ import annotations
+
+import argparse
+import gc
+import json
+import tempfile
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import torch
+from datasets import Dataset
+from peft import LoraConfig, TaskType, get_peft_model
+from trl import DPOTrainer, DPOConfig
+
+from wisent.comparison.utils import (
+    generate_contrastive_pairs,
+    create_test_only_task,
+    extract_accuracy,
+    run_lm_eval_evaluation,
+    run_ll_evaluation,
+    load_model_and_tokenizer,
+    apply_steering_to_model,
+    remove_steering,
+)
+from wisent.core.utils.device import preferred_dtype
+
+if TYPE_CHECKING:
+    from wisent.core.models.wisent_model import WisentModel
+
+
+def create_dpo_dataset(pairs: list[dict], tokenizer) -> Dataset:
+    """
+    Convert contrastive pairs to DPO dataset format.
+
+    - Chat models: returns conversational format, DPOTrainer auto-applies chat template
+    - Base models: returns text format with simple formatting
+
+    Args:
+        pairs: List of contrastive pairs
+        tokenizer: Tokenizer to check for chat template support
+
+    Returns:
+        HuggingFace Dataset ready for DPOTrainer
+    """
+    data = {
+        "prompt": [],
+        "chosen": [],
+        "rejected": [],
+    }
+
+    for pair in pairs:
+        prompt = pair["prompt"]
+        chosen = pair["positive_response"]["model_response"]
+        rejected = pair["negative_response"]["model_response"]
+
+        if hasattr(tokenizer, "chat_template") and tokenizer.chat_template:
+            # Chat model: use conversational format, trainer applies template
+            data["prompt"].append([{"role": "user", "content": prompt}])
+            data["chosen"].append([{"role": "assistant", "content": chosen}])
+            data["rejected"].append([{"role": "assistant", "content": rejected}])
+        else:
+            # Base model: use simple text format
+            data["prompt"].append(f"{prompt}\n")
+            data["chosen"].append(chosen)
+            data["rejected"].append(rejected)
+
+    return Dataset.from_dict(data)
+
+
+def train_lora_dpo(
+    task: str,
+    model_name: str,
+    output_path: str | Path,
+    num_pairs: int = 50,
+    device: str = "cuda:0",
+    keep_intermediate: bool = False,
+    lora_r: int = 16,
+    lora_alpha: int = 32,
+    lora_dropout: float = 0.05,
+    learning_rate: float = 5e-5,
+    num_epochs: int = 1,
+    batch_size: int = 1,
+    max_length: int = 512,
+    max_prompt_length: int = 256,
+    beta: float = 0.1,
+) -> Path:
+    """
+    Train a LoRA adapter using DPO on contrastive pairs from an lm-eval task.
+
+    Args:
+        task: lm-eval task name (e.g., 'boolq', 'cb')
+        model_name: HuggingFace model name
+        output_path: Where to save the trained adapter
+        num_pairs: Number of preference pairs to use
+        device: Device to run on
+        keep_intermediate: Whether to keep intermediate files
+        lora_r: LoRA rank
+        lora_alpha: LoRA alpha
+        lora_dropout: LoRA dropout
+        learning_rate: Learning rate
+        num_epochs: Number of training epochs
+        batch_size: Training batch size
+        max_length: Max total sequence length
+        max_prompt_length: Max prompt length
+        beta: DPO beta parameter (controls deviation from reference model)
+
+    Returns:
+        Path to saved adapter
+    """
+    output_path = Path(output_path)
+
+    # Step 1: Generate contrastive pairs
+    print(f"\n{'='*60}")
+    print(f"Step 1: Generating {num_pairs} preference pairs from {task}")
+    print(f"{'='*60}")
+
+    pairs, pairs_file = generate_contrastive_pairs(task, num_pairs)
+    print(f"Generated {len(pairs)} preference pairs")
+
+    # Step 2: Load model
+    print(f"\n{'='*60}")
+    print(f"Step 2: Loading model {model_name}")
+    print(f"{'='*60}")
+
+    model, tokenizer = load_model_and_tokenizer(model_name, device, eval_mode=False)
+
+    # Ensure tokenizer has padding
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.padding_side = "left"  # DPO typically uses left padding
+
+    # Step 3: Create DPO dataset
+    print(f"\n{'='*60}")
+    print(f"Step 3: Creating DPO dataset")
+    print(f"{'='*60}")
+
+    dataset = create_dpo_dataset(pairs, tokenizer)
+    print(f"Dataset size: {len(dataset)}")
+
+    # Step 4: Configure LoRA
+    print(f"\n{'='*60}")
+    print(f"Step 4: Configuring LoRA (r={lora_r}, alpha={lora_alpha})")
+    print(f"{'='*60}")
+
+    lora_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        r=lora_r,
+        lora_alpha=lora_alpha,
+        lora_dropout=lora_dropout,
+        target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
+        bias="none",
+    )
+
+    model = get_peft_model(model, lora_config)
+    model.print_trainable_parameters()
+
+    # Step 5: Configure DPO training
+    print(f"\n{'='*60}")
+    print(f"Step 5: Configuring DPO training")
+    print(f"{'='*60}")
+
+    training_output_dir = tempfile.mkdtemp(prefix="lora_dpo_training_")
+
+    # Determine dtype
+    dtype = preferred_dtype(device)
+
+    training_args = DPOConfig(
+        output_dir=training_output_dir,
+        num_train_epochs=num_epochs,
+        per_device_train_batch_size=batch_size,
+        gradient_accumulation_steps=1,
+        learning_rate=learning_rate,
+        weight_decay=0.01,
+        warmup_ratio=0.1,
+        logging_steps=10,
+        save_strategy="no",
+        bf16=(dtype == torch.bfloat16),
+        fp16=(dtype == torch.float16),
+        report_to="none",
+        max_length=max_length,
+        max_prompt_length=max_prompt_length,
+        beta=beta,
+        loss_type="sigmoid",  # Standard DPO loss
+    )
+
+    print(f"Beta: {beta}")
+    print(f"Max length: {max_length}")
+    print(f"Max prompt length: {max_prompt_length}")
+    print(f"Learning rate: {learning_rate}")
+    print(f"Epochs: {num_epochs}")
+    print(f"Batch size: {batch_size}")
+
+    # Step 6: Train with DPO
+    print(f"\n{'='*60}")
+    print(f"Step 6: Training with DPO")
+    print(f"{'='*60}")
+
+    trainer = DPOTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=dataset,
+        processing_class=tokenizer,
+    )
+
+    trainer.train()
+
+    # Step 7: Save adapter
+    print(f"\n{'='*60}")
+    print(f"Step 7: Saving LoRA adapter")
+    print(f"{'='*60}")
+
+    output_path.mkdir(parents=True, exist_ok=True)
+    model.save_pretrained(output_path)
+    tokenizer.save_pretrained(output_path)
+
+    # Save metadata
+    metadata = {
+        "task": task,
+        "model": model_name,
+        "training_method": "dpo",
+        "num_pairs": len(pairs),
+        "lora_r": lora_r,
+        "lora_alpha": lora_alpha,
+        "lora_dropout": lora_dropout,
+        "learning_rate": learning_rate,
+        "num_epochs": num_epochs,
+        "batch_size": batch_size,
+        "max_length": max_length,
+        "max_prompt_length": max_prompt_length,
+        "beta": beta,
+    }
+    with open(output_path / "metadata.json", "w") as f:
+        json.dump(metadata, f, indent=2)
+
+    # Cleanup
+    del model, trainer
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    if not keep_intermediate:
+        import os
+        import shutil
+        os.unlink(pairs_file)
+        shutil.rmtree(training_output_dir, ignore_errors=True)
+
+    print(f"\nDPO LoRA adapter saved to {output_path}")
+    return output_path
+
+
+def evaluate_lora_dpo(
+    model_name: str,
+    lora_path: str | Path,
+    task: str,
+    train_ratio: float = 0.8,
+    device: str = "cuda:0",
+    batch_size: int = 1,
+    max_batch_size: int = 8,
+    limit: int | None = None,
+    output_dir: str | Path = None,
+    # Training metadata (for output)
+    num_train_pairs: int | None = None,
+    num_epochs: int | None = None,
+    lora_r: int | None = None,
+    lora_alpha: int | None = None,
+    lora_dropout: float | None = None,
+    learning_rate: float | None = None,
+    beta: float | None = None,
+    max_length: int | None = None,
+    max_prompt_length: int | None = None,
+    # Steering parameters (optional)
+    with_steering: bool = False,
+    steering_method: str = "caa",
+    steering_layers: str = "12",
+    steering_num_pairs: int = 50,
+    steering_scales: list[float] | None = None,
+    extraction_strategy: str = "mc_completion",
+) -> dict:
+    """
+    Evaluate a trained DPO LoRA adapter.
+
+    Compares base model vs DPO-LoRA model accuracy.
+    Optionally also evaluates DPO-LoRA + steering at multiple scales.
+    """
+    from wisent.core.models.wisent_model import WisentModel
+    from wisent.comparison.lora import apply_lora_to_model, remove_lora
+
+    lora_path = Path(lora_path)
+
+    if steering_scales is None:
+        steering_scales = [1.0, 2.0, 4.0]
+
+    # Create test task
+    print(f"\n{'='*60}")
+    print(f"Creating test task for: {task}")
+    print(f"{'='*60}")
+
+    task_dict = create_test_only_task(task, train_ratio=train_ratio)
+
+    # Load model
+    print(f"\n{'='*60}")
+    print(f"Loading model: {model_name}")
+    print(f"{'='*60}")
+    wisent_model = WisentModel(model_name=model_name, device=device)
+
+    # Base evaluation
+    print(f"\n{'='*60}")
+    print(f"Running BASE evaluation")
+    print(f"{'='*60}")
+
+    base_results = run_lm_eval_evaluation(wisent_model, task_dict, task, batch_size, max_batch_size, limit)
+    base_acc_lm_eval = extract_accuracy(base_results, task)
+    print(f"Base accuracy (lm-eval): {base_acc_lm_eval:.4f}")
+
+    base_acc_ll = run_ll_evaluation(wisent_model, task_dict, task, limit)
+    print(f"Base accuracy (LL): {base_acc_ll:.4f}")
+
+    # Apply DPO LoRA
+    print(f"\n{'='*60}")
+    print(f"Applying DPO LoRA adapter from: {lora_path}")
+    print(f"{'='*60}")
+    apply_lora_to_model(wisent_model, lora_path)
+
+    # LoRA evaluation
+    print(f"\n{'='*60}")
+    print(f"Running DPO-LORA evaluation")
+    print(f"{'='*60}")
+
+    lora_results = run_lm_eval_evaluation(wisent_model, task_dict, task, batch_size, max_batch_size, limit)
+    lora_acc_lm_eval = extract_accuracy(lora_results, task)
+    print(f"DPO-LoRA accuracy (lm-eval): {lora_acc_lm_eval:.4f}")
+
+    lora_acc_ll = run_ll_evaluation(wisent_model, task_dict, task, limit)
+    print(f"DPO-LoRA accuracy (LL): {lora_acc_ll:.4f}")
+
+    # Results dict
+    results = {
+        "task": task,
+        "model": model_name,
+        "training_method": "dpo",
+        "lora_path": str(lora_path),
+        # Training config
+        "num_train_pairs": num_train_pairs,
+        "num_epochs": num_epochs,
+        "lora_r": lora_r,
+        "lora_alpha": lora_alpha,
+        "lora_dropout": lora_dropout,
+        "learning_rate": learning_rate,
+        "beta": beta,
+        "max_length": max_length,
+        "max_prompt_length": max_prompt_length,
+        # Eval config
+        "train_ratio": train_ratio,
+        "eval_limit": limit,
+        # Results
+        "base_accuracy_lm_eval": base_acc_lm_eval,
+        "base_accuracy_ll": base_acc_ll,
+        "lora_accuracy_lm_eval": lora_acc_lm_eval,
+        "lora_accuracy_ll": lora_acc_ll,
+        "lora_diff_lm_eval": lora_acc_lm_eval - base_acc_lm_eval,
+        "lora_diff_ll": lora_acc_ll - base_acc_ll,
+    }
+
+    # DPO-LoRA + Steering evaluation (if enabled)
+    if with_steering:
+        from wisent.core.trainers.steering_trainer import WisentSteeringTrainer
+        from wisent.core.steering_methods import get_steering_method
+        from wisent.core.activations.extraction_strategy import ExtractionStrategy
+        from wisent.core.contrastive_pairs.core.set import ContrastivePairSet
+        from wisent.core.contrastive_pairs.core.pair import ContrastivePair
+        from wisent.core.contrastive_pairs.core.response import PositiveResponse, NegativeResponse
+
+        # Generate contrastive pairs for steering
+        print(f"\n{'='*60}")
+        print(f"Generating {steering_num_pairs} contrastive pairs for steering")
+        print(f"{'='*60}")
+        pairs_data, pairs_file = generate_contrastive_pairs(task, steering_num_pairs)
+
+        # Convert to ContrastivePairSet
+        pairs = []
+        for p in pairs_data:
+            pair = ContrastivePair(
+                prompt=p["prompt"],
+                positive_response=PositiveResponse(model_response=p["positive_response"]["model_response"]),
+                negative_response=NegativeResponse(model_response=p["negative_response"]["model_response"]),
+            )
+            pairs.append(pair)
+        pair_set = ContrastivePairSet(pairs=pairs, name=f"{task}_dpo_lora_steering")
+        print(f"Created {len(pair_set)} contrastive pairs")
+
+        # Generate steering vector on DPO-LoRA model
+        print(f"\n{'='*60}")
+        print(f"Generating {steering_method.upper()} steering vector on DPO-LoRA model")
+        print(f"Layers: {steering_layers}")
+        print(f"{'='*60}")
+
+        steering_method_obj = get_steering_method(steering_method, device=device)
+        strategy = ExtractionStrategy(extraction_strategy)
+
+        trainer = WisentSteeringTrainer(
+            model=wisent_model,
+            pair_set=pair_set,
+            steering_method=steering_method_obj,
+        )
+
+        result = trainer.run(
+            layers_spec=steering_layers,
+            strategy=strategy,
+            accept_low_quality_vector=True,
+        )
+
+        # Convert to dict format for apply_steering_to_model
+        steering_vectors = {}
+        for layer_name, tensor in result.steered_vectors.to_dict().items():
+            if tensor is not None:
+                steering_vectors[layer_name] = tensor.cpu().float().tolist()
+
+        steering_data = {
+            "steering_vectors": steering_vectors,
+            "layers": list(steering_vectors.keys()),
+        }
+
+        # Cleanup temp file
+        import os
+        os.unlink(pairs_file)
+
+        # Add steering info to results
+        results["steering"] = {
+            "method": steering_method,
+            "layers": list(steering_vectors.keys()),
+            "num_pairs": steering_num_pairs,
+            "extraction_strategy": extraction_strategy,
+            "scales": {},
+        }
+
+        # Evaluate at each scale
+        for scale in steering_scales:
+            print(f"\n{'='*60}")
+            print(f"Evaluating DPO-LoRA+{steering_method.upper()} at scale={scale}")
+            print(f"{'='*60}")
+
+            apply_steering_to_model(wisent_model, steering_data, scale=scale)
+
+            steer_results = run_lm_eval_evaluation(wisent_model, task_dict, task, batch_size, max_batch_size, limit)
+            steer_acc_lm_eval = extract_accuracy(steer_results, task)
+            print(f"DPO-LoRA+{steering_method.upper()} accuracy (lm-eval): {steer_acc_lm_eval:.4f}")
+
+            steer_acc_ll = run_ll_evaluation(wisent_model, task_dict, task, limit)
+            print(f"DPO-LoRA+{steering_method.upper()} accuracy (LL): {steer_acc_ll:.4f}")
+
+            remove_steering(wisent_model)
+
+            results["steering"]["scales"][str(scale)] = {
+                "accuracy_lm_eval": steer_acc_lm_eval,
+                "accuracy_ll": steer_acc_ll,
+                "diff_from_base_lm_eval": steer_acc_lm_eval - base_acc_lm_eval,
+                "diff_from_base_ll": steer_acc_ll - base_acc_ll,
+                "diff_from_lora_lm_eval": steer_acc_lm_eval - lora_acc_lm_eval,
+                "diff_from_lora_ll": steer_acc_ll - lora_acc_ll,
+            }
+
+    # Cleanup
+    remove_lora(wisent_model)
+    del wisent_model
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    # Print summary
+    print(f"\n{'='*70}")
+    print(f"RESULTS SUMMARY")
+    print(f"{'='*70}")
+    print(f"Task: {task}")
+    print(f"Model: {model_name}")
+    print(f"Training: DPO")
+    print(f"{'-'*70}")
+    print(f"{'Method':<25} {'lm-eval acc':<15} {'LL acc':<15} {'Diff (lm-eval)':<15}")
+    print(f"{'-'*70}")
+    print(f"{'Base':<25} {base_acc_lm_eval:<15.4f} {base_acc_ll:<15.4f} {'':<15}")
+    print(f"{'DPO-LoRA':<25} {lora_acc_lm_eval:<15.4f} {lora_acc_ll:<15.4f} {lora_acc_lm_eval - base_acc_lm_eval:+.4f}")
+
+    if with_steering:
+        for scale, res in results["steering"]["scales"].items():
+            label = f"DPO-LoRA+{steering_method.upper()}@{scale}"
+            print(f"{label:<25} {res['accuracy_lm_eval']:<15.4f} {res['accuracy_ll']:<15.4f} {res['diff_from_base_lm_eval']:+.4f}")
+
+    print(f"{'='*70}")
+
+    # Save results
+    if output_dir:
+        output_dir = Path(output_dir)
+        model_dir_name = model_name.replace("/", "_")
+        output_dir = output_dir / model_dir_name
+        output_dir.mkdir(parents=True, exist_ok=True)
+        results_file = output_dir / f"{task}_lora_dpo_eval_results.json"
+        with open(results_file, "w") as f:
+            json.dump(results, f, indent=2)
+        print(f"\nResults saved to: {results_file}")
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Train and evaluate LoRA adapter using DPO")
+    parser.add_argument("--model", required=True, help="HuggingFace model name")
+    parser.add_argument("--task", default="boolq", help="lm-eval task name")
+    parser.add_argument("--output-dir", default="/home/ubuntu/output", help="Output directory")
+    parser.add_argument("--num-pairs", type=int, default=50, help="Number of preference pairs")
+    parser.add_argument("--device", default="cuda:0", help="Device")
+    parser.add_argument("--lora-r", type=int, default=16, help="LoRA rank")
+    parser.add_argument("--lora-alpha", type=int, default=32, help="LoRA alpha")
+    parser.add_argument("--lora-dropout", type=float, default=0.05, help="LoRA dropout")
+    parser.add_argument("--learning-rate", type=float, default=5e-5, help="Learning rate")
+    parser.add_argument("--num-epochs", type=int, default=1, help="Number of epochs")
+    parser.add_argument("--batch-size", type=int, default=1, help="Training batch size")
+    parser.add_argument("--max-length", type=int, default=512, help="Max total sequence length")
+    parser.add_argument("--max-prompt-length", type=int, default=256, help="Max prompt length")
+    parser.add_argument("--beta", type=float, default=0.1, help="DPO beta (controls KL penalty)")
+    parser.add_argument("--keep-intermediate", action="store_true", help="Keep intermediate files")
+    # Eval args
+    parser.add_argument("--train-ratio", type=float, default=0.8, help="Train/test split ratio")
+    parser.add_argument("--eval-batch-size", default="auto", help="Eval batch size")
+    parser.add_argument("--eval-max-batch-size", type=int, default=64, help="Max eval batch size")
+    parser.add_argument("--eval-limit", type=int, default=None, help="Limit eval examples")
+    parser.add_argument("--skip-eval", action="store_true", help="Skip evaluation after training")
+    # DPO-LoRA + Steering args
+    parser.add_argument("--with-steering", action="store_true", help="Also evaluate DPO-LoRA + steering")
+    parser.add_argument("--steering-method", default="caa", choices=["caa", "fgaa"], help="Steering method")
+    parser.add_argument("--steering-layers", default="12", help="Layers for steering vector")
+    parser.add_argument("--steering-num-pairs", type=int, default=50, help="Number of pairs for steering")
+    parser.add_argument("--steering-scales", default="1.0,2.0,4.0", help="Comma-separated steering scales")
+    parser.add_argument("--extraction-strategy", default="mc_balanced", help="Extraction strategy for steering")
+
+    args = parser.parse_args()
+
+    output_path = Path(args.output_dir) / f"{args.task}_lora_dpo_adapter"
+
+    # Train
+    train_lora_dpo(
+        task=args.task,
+        model_name=args.model,
+        output_path=output_path,
+        num_pairs=args.num_pairs,
+        device=args.device,
+        keep_intermediate=args.keep_intermediate,
+        lora_r=args.lora_r,
+        lora_alpha=args.lora_alpha,
+        lora_dropout=args.lora_dropout,
+        learning_rate=args.learning_rate,
+        num_epochs=args.num_epochs,
+        batch_size=args.batch_size,
+        max_length=args.max_length,
+        max_prompt_length=args.max_prompt_length,
+        beta=args.beta,
+    )
+
+    # Evaluate
+    if not args.skip_eval:
+        eval_batch_size = args.eval_batch_size
+        if eval_batch_size != "auto":
+            eval_batch_size = int(eval_batch_size)
+
+        # Parse steering scales
+        steering_scales = [float(s.strip()) for s in args.steering_scales.split(",")]
+
+        evaluate_lora_dpo(
+            model_name=args.model,
+            lora_path=output_path,
+            task=args.task,
+            train_ratio=args.train_ratio,
+            device=args.device,
+            batch_size=eval_batch_size,
+            max_batch_size=args.eval_max_batch_size,
+            limit=args.eval_limit,
+            output_dir=args.output_dir,
+            # Training metadata
+            num_train_pairs=args.num_pairs,
+            num_epochs=args.num_epochs,
+            lora_r=args.lora_r,
+            lora_alpha=args.lora_alpha,
+            lora_dropout=args.lora_dropout,
+            learning_rate=args.learning_rate,
+            beta=args.beta,
+            max_length=args.max_length,
+            max_prompt_length=args.max_prompt_length,
+            # Steering parameters
+            with_steering=args.with_steering,
+            steering_method=args.steering_method,
+            steering_layers=args.steering_layers,
+            steering_num_pairs=args.steering_num_pairs,
+            steering_scales=steering_scales,
+            extraction_strategy=args.extraction_strategy,
+        )
+
+
+if __name__ == "__main__":
+    main()