lightonocr-finetune.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "torch>=2.0.0",
#     "datasets>=2.14.0",
#     "accelerate>=0.24.0",
#     "huggingface-hub",
#     "pillow>=12.0.0",
#     "jiwer>=3.0.0",
#     "tqdm>=4.65.0",
#     "transformers @ git+https://github.com/baptiste-aubertin/transformers.git@main",
# ]
# ///

"""
Fine-tune LightOnOCR on OCR datasets.

LightOnOCR is an end-to-end trainable vision-language model specifically designed for OCR tasks.
This script enables fine-tuning on custom datasets for improved performance on specific domains,
languages, or document types.

Examples:
    # Basic fine-tuning on IAM handwriting dataset
    uv run lightonocr-finetune.py \
        --dataset-id HuggingFaceM4/FineVision \
        --subset iam \
        --output-dir ./lightonocr-iam \
        --epochs 2

    # Fine-tune with frozen language model to save memory
    uv run lightonocr-finetune.py \
        --dataset-id HuggingFaceM4/FineVision \
        --subset olmOCR-mix-0225-documents \
        --output-dir ./lightonocr-docs \
        --freeze-language \
        --batch-size 8
    
    # Stream large datasets to reduce memory usage
    uv run lightonocr-finetune.py \
        --dataset-id HuggingFaceM4/FineVision \
        --subset olmOCR-mix-0225-books \
        --output-dir ./lightonocr-books \
        --streaming \
        --shuffle-buffer-size 10000 \
        --max-train-samples 5000  # Will auto-calculate max-steps

    # Push to Hub with evaluation metrics
    uv run lightonocr-finetune.py \
        --dataset-id HuggingFaceM4/FineVision \
        --subset iam \
        --hub-model-id username/lightonocr-iam \
        --push-to-hub \
        --eval-samples 100

    # Run on HF Jobs with GPU and streaming
    hf jobs run --gpu l4x1 \
        uv run lightonocr-finetune.py \
        --dataset-id custom/large-ocr-dataset \
        --output-dir ./custom-ocr \
        --streaming \
        --epochs 3
"""

import argparse
import logging
import os
import sys
from datetime import datetime
from pathlib import Path
from typing import Dict, Optional

import torch
from datasets import load_dataset
from huggingface_hub import login
from jiwer import cer, wer
from tqdm import tqdm
from transformers import (
    AutoProcessor,
    LightOnOCRForConditionalGeneration,
    Trainer,
    TrainingArguments,
)

logging.basicConfig(
    level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)

os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"

# Constants for the assistant pattern in chat template
ASSISTANT_START_PATTERN = [151645, 1699, 151644, 77091, 1699]
DEFAULT_MAX_LENGTH = 1024
DEFAULT_LONGEST_EDGE = 700


class OCRDataCollator:
    """Data collator for OCR fine-tuning."""

    def __init__(
        self,
        processor,
        max_length=DEFAULT_MAX_LENGTH,
        longest_edge=DEFAULT_LONGEST_EDGE,
    ):
        self.processor = processor
        self.max_length = max_length
        self.longest_edge = longest_edge

    def __call__(self, examples):
        batch_messages = []
        batch_images = []

        for example in examples:
            example_images = example["images"]
            example_texts = example["texts"]

            # Validate single image/text per example
            if len(example_images) != 1 or len(example_texts) != 1:
                logger.warning(
                    f"Skipping example with {len(example_images)} images and {len(example_texts)} texts"
                )
                continue

            image = example_images[0].convert("RGB")
            batch_images.append(image)

            # Extract assistant text from conversation
            conversation = example_texts[0]
            assistant_text = conversation.get("assistant", "").strip()

            messages = [
                {"role": "user", "content": [{"type": "image"}]},
                {
                    "role": "assistant",
                    "content": [{"type": "text", "text": assistant_text}],
                },
            ]
            batch_messages.append(messages)

        if not batch_images:
            return None

        # Apply chat template
        texts = [
            self.processor.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=False
            )
            for messages in batch_messages
        ]

        # Process inputs
        inputs = self.processor(
            text=texts,
            images=batch_images,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=self.max_length,
            size={"longest_edge": self.longest_edge},
        )

        # Create labels (mask prompt, train only on assistant response)
        labels = inputs["input_ids"].clone()
        pad_token_id = self.processor.tokenizer.pad_token_id

        for i in range(len(labels)):
            full_ids = inputs["input_ids"][i].tolist()

            # Find where assistant content starts
            assistant_content_start = None

            # Try the standard pattern: <|im_end|>\n<|im_start|>assistant\n
            for idx in range(len(full_ids) - len(ASSISTANT_START_PATTERN)):
                if (
                    full_ids[idx : idx + len(ASSISTANT_START_PATTERN)]
                    == ASSISTANT_START_PATTERN
                ):
                    assistant_content_start = idx + len(ASSISTANT_START_PATTERN)
                    break

            if assistant_content_start is None:
                # Some samples may not have the exact pattern - this is expected
                # The model will train on samples where the pattern is found
                labels[i, :] = -100
            else:
                # Mask everything first
                labels[i, :] = -100

                # Unmask from assistant content start to end
                for idx in range(assistant_content_start, len(full_ids)):
                    if full_ids[idx] == pad_token_id:
                        break
                    labels[i, idx] = inputs["input_ids"][i, idx]

            # Mask padding tokens
            labels[i, inputs["input_ids"][i] == pad_token_id] = -100

        inputs["labels"] = labels

        # Convert to proper dtype
        inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

        return inputs


def evaluate_model(
    model,
    processor,
    dataset,
    num_samples: int = 50,
    batch_size: int = 8,
    device: str = "cuda",
    description: str = "Model",
    is_streaming: bool = False,
) -> Dict[str, float]:
    """
    Evaluate model on dataset and compute OCR metrics.

    Args:
        model: The model to evaluate
        processor: The processor for the model
        dataset: Dataset to evaluate on (can be streaming or regular)
        num_samples: Number of samples to evaluate
        batch_size: Batch size for evaluation
        device: Device to run evaluation on
        description: Description for logging
        is_streaming: Whether the dataset is a streaming dataset

    Returns:
        Dictionary with CER, WER, and perfect match count
    """
    model.eval()
    predictions = []
    ground_truths = []

    logger.info(f"Evaluating {description} on {num_samples} samples...")

    # Handle streaming datasets differently
    if is_streaming:
        # For streaming datasets, we take the first num_samples
        samples_processed = 0
        batch_samples = []

        for sample in tqdm(dataset, total=num_samples, desc="Evaluating"):
            if samples_processed >= num_samples:
                break

            batch_samples.append(sample)
            samples_processed += 1

            # Process when we have a full batch or reached the end
            if len(batch_samples) == batch_size or samples_processed == num_samples:
                batch_images = [[s["images"][0]] for s in batch_samples]
                batch_ground_truths = [
                    s["texts"][0]["assistant"].strip() for s in batch_samples
                ]

                # Prepare inputs
                messages = [{"role": "user", "content": [{"type": "image"}]}]
                text = processor.apply_chat_template(
                    messages, tokenize=False, add_generation_prompt=True
                )
                texts = [text] * len(batch_images)

                inputs = processor(
                    text=texts,
                    images=batch_images,
                    return_tensors="pt",
                    padding=True,
                    truncation=True,
                    max_length=DEFAULT_MAX_LENGTH,
                    size={"longest_edge": DEFAULT_LONGEST_EDGE},
                ).to(device)
                inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

                # Generate predictions
                with torch.no_grad():
                    outputs = model.generate(
                        **inputs, max_new_tokens=512, do_sample=True
                    )

                input_length = inputs["input_ids"].shape[1]
                generated_ids = outputs[:, input_length:]
                batch_predictions = processor.batch_decode(
                    generated_ids, skip_special_tokens=True
                )
                batch_predictions = [p.strip() for p in batch_predictions]

                predictions.extend(batch_predictions)
                ground_truths.extend(batch_ground_truths)
                batch_samples = []
    else:
        # Original non-streaming evaluation
        for start_idx in tqdm(range(0, min(num_samples, len(dataset)), batch_size)):
            end_idx = min(start_idx + batch_size, num_samples, len(dataset))
            batch_samples = [dataset[i] for i in range(start_idx, end_idx)]

            batch_images = [[s["images"][0]] for s in batch_samples]
            batch_ground_truths = [
                s["texts"][0]["assistant"].strip() for s in batch_samples
            ]

            # Prepare inputs
            messages = [{"role": "user", "content": [{"type": "image"}]}]
            text = processor.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True
            )
            texts = [text] * len(batch_images)

            inputs = processor(
                text=texts,
                images=batch_images,
                return_tensors="pt",
                padding=True,
                truncation=True,
                max_length=DEFAULT_MAX_LENGTH,
                size={"longest_edge": DEFAULT_LONGEST_EDGE},
            ).to(device)
            inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

            # Generate predictions
            with torch.no_grad():
                outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True)

            input_length = inputs["input_ids"].shape[1]
            generated_ids = outputs[:, input_length:]
            batch_predictions = processor.batch_decode(
                generated_ids, skip_special_tokens=True
            )
            batch_predictions = [p.strip() for p in batch_predictions]

            predictions.extend(batch_predictions)
            ground_truths.extend(batch_ground_truths)

    # Compute metrics
    cer_score = cer(ground_truths, predictions) * 100
    wer_score = wer(ground_truths, predictions) * 100
    perfect_matches = sum(
        1 for pred, gt in zip(predictions, ground_truths) if pred == gt
    )

    actual_samples = len(predictions)
    logger.info(
        f"CER: {cer_score:.2f}% | WER: {wer_score:.2f}% | Perfect: {perfect_matches}/{actual_samples}"
    )

    # Show a few examples
    for i in range(min(3, len(predictions))):
        match = "✅" if predictions[i] == ground_truths[i] else "❌"
        logger.info(
            f"{match} Sample {i + 1}: '{predictions[i][:50]}...' vs '{ground_truths[i][:50]}...'"
        )

    return {
        "cer": cer_score,
        "wer": wer_score,
        "perfect_matches": perfect_matches,
        "total_samples": actual_samples,
    }


def create_model_card_content(
    model_id: str,
    dataset_id: str,
    subset: Optional[str],
    base_metrics: Dict[str, float],
    finetuned_metrics: Dict[str, float],
    training_args: TrainingArguments,
    freeze_config: Dict[str, bool],
) -> str:
    """Generate model card content with training details and metrics."""

    # Calculate improvements
    cer_improvement = base_metrics["cer"] - finetuned_metrics["cer"]
    wer_improvement = base_metrics["wer"] - finetuned_metrics["wer"]
    perfect_improvement = (
        finetuned_metrics["perfect_matches"] - base_metrics["perfect_matches"]
    )

    # Determine which components were frozen
    frozen_components = [comp for comp, is_frozen in freeze_config.items() if is_frozen]
    frozen_str = (
        ", ".join(frozen_components) if frozen_components else "None (full fine-tuning)"
    )

    dataset_str = f"{dataset_id}/{subset}" if subset else dataset_id

    content = f"""---
license: mit
tags:
- vision
- ocr
- document-understanding
- transformers
base_model: lightonai/LightOnOCR-1B-1025
datasets:
- {dataset_id}
metrics:
- cer
- wer
library_name: transformers
---

# {model_id.split("/")[-1]}

This model is a fine-tuned version of [LightOnOCR-1B-1025](https://huggingface.co/lightonai/LightOnOCR-1B-1025) on the {dataset_str} dataset.

## Model Description

LightOnOCR is an end-to-end trainable vision-language model specifically designed for OCR tasks. This fine-tuned version has been optimized for improved performance on the target dataset.

## Training Details

### Dataset
- **Source**: {dataset_str}
- **Training samples**: {training_args.num_train_epochs * (training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps)}
- **Validation samples**: Used for model selection

### Training Configuration
- **Epochs**: {training_args.num_train_epochs}
- **Batch size**: {training_args.per_device_train_batch_size} (effective: {training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps})
- **Learning rate**: {training_args.learning_rate}
- **Frozen components**: {frozen_str}
- **Hardware**: GPU with mixed precision (bf16)

## Evaluation Results

### Performance Comparison

| Metric | Base Model | Fine-tuned | Improvement |
|--------|------------|------------|-------------|
| **CER (%)** | {base_metrics["cer"]:.2f} | {finetuned_metrics["cer"]:.2f} | {cer_improvement:+.2f} |
| **WER (%)** | {base_metrics["wer"]:.2f} | {finetuned_metrics["wer"]:.2f} | {wer_improvement:+.2f} |
| **Perfect Matches** | {base_metrics["perfect_matches"]}/{base_metrics["total_samples"]} | {finetuned_metrics["perfect_matches"]}/{finetuned_metrics["total_samples"]} | {perfect_improvement:+d} |

*Lower is better for CER and WER. Evaluation performed on {finetuned_metrics["total_samples"]} test samples.*

## Usage

```python
from transformers import AutoProcessor, LightOnOCRForConditionalGeneration
from PIL import Image
import torch

# Load model and processor
model = LightOnOCRForConditionalGeneration.from_pretrained(
    "{model_id}",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("{model_id}")

# Prepare image
image = Image.open("your_image.jpg").convert("RGB")

# Create prompt
messages = [
    {{"role": "user", "content": [{{"type": "image"}}]}}
]

# Process and generate
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(
    text=[text],
    images=[[image]],
    return_tensors="pt",
    max_length=1024
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
generated_text = processor.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(generated_text)
```

## Training Script

This model was trained using the UV Scripts training pipeline. To reproduce or further fine-tune:

```bash
uv run https://huggingface.co/datasets/uv-scripts/transformers-training/raw/main/lightonocr-finetune.py \\
    --dataset-id {dataset_id} \\
    {"--subset " + subset if subset else ""} \\
    --output-dir ./model \\
    --epochs {training_args.num_train_epochs}
```

## Citation

If you use this model, please cite:

```bibtex
@misc{{lightonocr2024,
  title={{LightOnOCR: End-to-End Trainable OCR Model}},
  author={{LightOn AI}},
  year={{2024}},
  url={{https://huggingface.co/blog/lightonai/lightonocr}}
}}
```

## License

This model is released under the MIT license.

---

*Generated on {datetime.now().strftime("%Y-%m-%d")} using [UV Scripts](https://huggingface.co/uv-scripts)*
"""

    return content


def main():
    parser = argparse.ArgumentParser(
        description="Fine-tune LightOnOCR on OCR datasets",
        formatter_class=argparse.RawDescriptionHelpFormatter,
    )

    # Dataset arguments
    parser.add_argument(
        "--dataset-id",
        type=str,
        default="HuggingFaceM4/FineVision",
        help="HuggingFace dataset ID",
    )
    parser.add_argument(
        "--subset",
        type=str,
        default="iam",
        choices=["iam", "olmOCR-mix-0225-books", "olmOCR-mix-0225-documents"],
        help="Dataset subset to use (for FineVision)",
    )
    parser.add_argument(
        "--train-split",
        type=str,
        default="train[:85%]",
        help="Training split specification",
    )
    parser.add_argument(
        "--val-split",
        type=str,
        default="train[85%:95%]",
        help="Validation split specification",
    )
    parser.add_argument(
        "--test-split", type=str, default="train[95%:]", help="Test split specification"
    )

    # Streaming arguments
    parser.add_argument(
        "--streaming",
        action="store_true",
        help="Use dataset streaming to reduce memory usage (Note: uses full training set, ignores train-split percentages)",
    )
    parser.add_argument(
        "--shuffle-buffer-size",
        type=int,
        default=10000,
        help="Buffer size for shuffling when using streaming (default: 10000)",
    )
    parser.add_argument(
        "--max-train-samples",
        type=int,
        help="Maximum number of training samples when streaming (useful for quick experiments)",
    )

    # Model arguments
    parser.add_argument(
        "--model-id",
        type=str,
        default="lightonai/LightOnOCR-1B-1025",
        help="Base model ID",
    )
    parser.add_argument(
        "--freeze-vision", action="store_true", help="Freeze vision encoder"
    )
    parser.add_argument(
        "--freeze-language", action="store_true", help="Freeze language model"
    )
    parser.add_argument(
        "--freeze-projection",
        action="store_true",
        help="Freeze vision projection layer",
    )

    # Training arguments
    parser.add_argument(
        "--output-dir", type=str, required=True, help="Directory to save the model"
    )
    parser.add_argument(
        "--epochs", type=int, default=2, help="Number of training epochs"
    )
    parser.add_argument(
        "--batch-size", type=int, default=4, help="Training batch size per device"
    )
    parser.add_argument(
        "--gradient-accumulation",
        type=int,
        default=4,
        help="Gradient accumulation steps",
    )
    parser.add_argument(
        "--learning-rate", type=float, default=6e-5, help="Learning rate"
    )
    parser.add_argument(
        "--warmup-steps", type=int, default=10, help="Number of warmup steps"
    )
    parser.add_argument(
        "--eval-steps", type=int, default=50, help="Evaluation interval (in steps)"
    )
    parser.add_argument(
        "--save-steps",
        type=int,
        default=500,
        help="Save checkpoint interval (in steps)",
    )
    parser.add_argument(
        "--max-length", type=int, default=1024, help="Maximum sequence length"
    )
    parser.add_argument(
        "--longest-edge", type=int, default=700, help="Longest edge for image resizing"
    )

    # Evaluation arguments
    parser.add_argument(
        "--eval-samples", type=int, default=100, help="Number of samples for evaluation"
    )
    parser.add_argument(
        "--eval-batch-size", type=int, default=8, help="Batch size for evaluation"
    )
    parser.add_argument(
        "--skip-base-eval", action="store_true", help="Skip base model evaluation"
    )

    # Hub arguments
    parser.add_argument(
        "--hub-model-id", type=str, help="HuggingFace Hub model ID for pushing"
    )
    parser.add_argument(
        "--push-to-hub", action="store_true", help="Push model to HuggingFace Hub"
    )
    parser.add_argument("--hf-token", type=str, help="HuggingFace API token")
    parser.add_argument(
        "--private", action="store_true", help="Make the model private on Hub"
    )

    # Other arguments
    parser.add_argument(
        "--max-samples", type=int, help="Limit number of training samples (for testing)"
    )
    parser.add_argument("--seed", type=int, default=42, help="Random seed")
    parser.add_argument(
        "--max-steps",
        type=int,
        help="Maximum number of training steps (auto-calculated for streaming if not specified)"
    )

    args = parser.parse_args()

    # Check GPU availability
    if not torch.cuda.is_available():
        logger.error("CUDA is not available. This script requires a GPU.")
        logger.info("To run on HF Jobs with GPU:")
        logger.info(
            f"hf jobs run --gpu l4x1 uv run {__file__} --dataset-id {args.dataset_id} --output-dir {args.output_dir}"
        )
        sys.exit(1)

    device = "cuda"
    logger.info(f"Using GPU: {torch.cuda.get_device_name(0)}")

    # Set environment variables for better performance
    os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
    os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
    torch.set_float32_matmul_precision("high")

    # Login to HuggingFace if needed
    if args.push_to_hub:
        token = args.hf_token or os.environ.get("HF_TOKEN")
        if token:
            login(token=token)
        else:
            logger.error("HF_TOKEN required for push_to_hub")
            sys.exit(1)

    # Load dataset
    logger.info(f"Loading dataset: {args.dataset_id}/{args.subset}")

    if args.streaming:
        logger.info("Using streaming mode for dataset loading")
        # For streaming, we can only use "train" split, not percentage-based splits
        # Load the full training set in streaming mode
        train_ds = load_dataset(
            args.dataset_id, args.subset, split="train", streaming=True
        )

        # For validation and test, we need to load a subset of the data
        # We'll use the last 15% of the data for validation and test
        # Load the full dataset for splitting into val/test
        full_ds = load_dataset(args.dataset_id, args.subset, split="train")
        total_size = len(full_ds)

        # Calculate split indices
        train_end = int(0.85 * total_size)
        val_end = int(0.95 * total_size)

        # Create validation and test splits
        val_ds = full_ds.select(range(train_end, val_end))
        test_ds = full_ds.select(range(val_end, total_size))

        # Clean up the full dataset to save memory
        del full_ds

        # Apply shuffling with buffer for streaming dataset
        train_ds = train_ds.shuffle(
            seed=args.seed, buffer_size=args.shuffle_buffer_size
        )

        # Limit samples if requested (for streaming)
        if args.max_samples or args.max_train_samples:
            max_samples = args.max_samples or args.max_train_samples
            train_ds = train_ds.take(max_samples)
            logger.info(f"Limited training to {max_samples} samples (streaming mode)")

        logger.info(
            f"Dataset loaded - Training: streaming (full train set), Val: {len(val_ds)}, Test: {len(test_ds)}"
        )
        logger.info(
            "Note: When streaming, using full training set. Use --max-train-samples to limit."
        )
    else:
        # Original non-streaming loading
        train_ds = load_dataset(args.dataset_id, args.subset, split=args.train_split)
        val_ds = load_dataset(args.dataset_id, args.subset, split=args.val_split)
        test_ds = load_dataset(args.dataset_id, args.subset, split=args.test_split)

        # Limit samples if requested (non-streaming)
        if args.max_samples:
            train_ds = train_ds.select(range(min(args.max_samples, len(train_ds))))
            logger.info(f"Limited training to {len(train_ds)} samples")

        logger.info(
            f"Dataset sizes - Train: {len(train_ds)}, Val: {len(val_ds)}, Test: {len(test_ds)}"
        )

    # Load processor
    logger.info(f"Loading processor from {args.model_id}")
    processor = AutoProcessor.from_pretrained(args.model_id)
    processor.tokenizer.padding_side = "left"

    # Load model
    logger.info(f"Loading model from {args.model_id}")
    model = LightOnOCRForConditionalGeneration.from_pretrained(
        args.model_id,
        torch_dtype=torch.bfloat16,
        attn_implementation="sdpa",
        device_map="auto",
    ).to(device)

    # Freeze components as requested
    freeze_config = {
        "vision_encoder": args.freeze_vision,
        "language_model": args.freeze_language,
        "vision_projection": args.freeze_projection,
    }

    if args.freeze_vision:
        for param in model.model.vision_encoder.parameters():
            param.requires_grad = False
        logger.info("Vision encoder frozen")

    if args.freeze_language:
        for param in model.model.language_model.parameters():
            param.requires_grad = False
        logger.info("Language model frozen")

    if args.freeze_projection:
        for param in model.model.vision_projection.parameters():
            param.requires_grad = False
        logger.info("Vision projection frozen")

    # Count trainable parameters
    total_params = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    logger.info(f"Total parameters: {total_params:,}")
    logger.info(
        f"Trainable parameters: {trainable_params:,} ({100 * trainable_params / total_params:.2f}%)"
    )

    # Evaluate base model
    base_metrics = {
        "cer": 0.0,
        "wer": 0.0,
        "perfect_matches": 0,
        "total_samples": args.eval_samples,
    }
    if not args.skip_base_eval:
        logger.info("\n" + "=" * 80)
        logger.info("EVALUATING BASE MODEL")
        logger.info("=" * 80)
        base_metrics = evaluate_model(
            model,
            processor,
            test_ds,
            num_samples=args.eval_samples,
            batch_size=args.eval_batch_size,
            device=device,
            description="Base model",
            is_streaming=False,  # Test dataset is never streamed
        )
        torch.cuda.empty_cache()

    # Prepare data collator
    data_collator = OCRDataCollator(
        processor, max_length=args.max_length, longest_edge=args.longest_edge
    )

    # Calculate max_steps for streaming datasets
    max_steps = None
    if args.streaming:
        if args.max_steps:
            max_steps = args.max_steps
            logger.info(f"Using user-specified max_steps: {max_steps}")
        else:
            # Estimate max_steps based on dataset size and batch configuration
            if args.max_train_samples:
                # Calculate based on limited samples
                steps_per_epoch = args.max_train_samples // (args.batch_size * args.gradient_accumulation)
                max_steps = steps_per_epoch * args.epochs
                logger.info(f"Calculated max_steps from max_train_samples: {max_steps}")
            else:
                # Use a default reasonable value
                # Approximate based on typical dataset sizes
                # Default to 1000 steps per epoch as a reasonable estimate
                max_steps = 1000 * args.epochs
                logger.warning(
                    f"Streaming mode: Using default max_steps={max_steps}. "
                    f"Consider setting --max-steps or --max-train-samples for precise control."
                )

    # Setup training arguments
    # When streaming, use max_steps instead of num_train_epochs
    training_args_dict = {
        "output_dir": args.output_dir,
        "per_device_train_batch_size": args.batch_size,
        "per_device_eval_batch_size": args.eval_batch_size,
        "gradient_accumulation_steps": args.gradient_accumulation,
        "learning_rate": args.learning_rate,
        "weight_decay": 0.0,
        "logging_steps": 50,
        "eval_strategy": "steps",
        "eval_steps": args.eval_steps,
        "save_strategy": "steps",
        "save_steps": args.save_steps,
        "save_total_limit": 2,
        "load_best_model_at_end": True,
        "metric_for_best_model": "eval_loss",
        "bf16": True,
        "fp16": False,
        "remove_unused_columns": False,
        "dataloader_pin_memory": False,
        "gradient_checkpointing": True,
        "optim": "adamw_torch_fused" if torch.cuda.is_available() else "adamw_torch",
        "warmup_steps": args.warmup_steps,
        "lr_scheduler_type": "linear",
        "push_to_hub": args.push_to_hub,
        "hub_model_id": args.hub_model_id,
        "hub_private_repo": args.private,
    }
    
    # Add either max_steps or num_train_epochs based on streaming mode
    if args.streaming:
        training_args_dict["max_steps"] = max_steps
        # Still set num_train_epochs for model card generation
        training_args_dict["num_train_epochs"] = args.epochs
    else:
        training_args_dict["num_train_epochs"] = args.epochs
    
    training_args = TrainingArguments(**training_args_dict)

    # Use smaller validation set for faster evaluation
    val_ds_small = val_ds.select(range(min(100, len(val_ds))))

    # Create trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_ds,
        eval_dataset=val_ds_small,
        data_collator=data_collator,
    )

    # Train
    logger.info("\n" + "=" * 80)
    logger.info("STARTING TRAINING")
    logger.info("=" * 80)
    if args.streaming:
        logger.info(
            f"Training samples: streaming mode (max: {args.max_train_samples or 'unlimited'})"
        )
        logger.info(f"Max steps: {max_steps}")
    else:
        logger.info(f"Training samples: {len(train_ds)}")
    logger.info(f"Validation samples: {len(val_ds_small)}")
    logger.info(f"Effective batch size: {args.batch_size * args.gradient_accumulation}")

    trainer.train()

    # Save model
    logger.info("Saving model and processor...")
    trainer.save_model(args.output_dir)
    processor.save_pretrained(args.output_dir)

    # Evaluate fine-tuned model
    logger.info("\n" + "=" * 80)
    logger.info("EVALUATING FINE-TUNED MODEL")
    logger.info("=" * 80)
    finetuned_metrics = evaluate_model(
        model,
        processor,
        test_ds,
        num_samples=args.eval_samples,
        batch_size=args.eval_batch_size,
        device=device,
        description="Fine-tuned model",
        is_streaming=False,  # Test dataset is never streamed
    )

    # Show comparison
    if not args.skip_base_eval:
        logger.info("\n" + "=" * 80)
        logger.info("PERFORMANCE COMPARISON")
        logger.info("=" * 80)
        logger.info(
            f"{'Metric':<20} {'Base':<12} {'Fine-tuned':<12} {'Improvement':<12}"
        )
        logger.info("-" * 56)
        logger.info(
            f"{'CER (%)':<20} {base_metrics['cer']:<12.2f} {finetuned_metrics['cer']:<12.2f} {base_metrics['cer'] - finetuned_metrics['cer']:+.2f}"
        )
        logger.info(
            f"{'WER (%)':<20} {base_metrics['wer']:<12.2f} {finetuned_metrics['wer']:<12.2f} {base_metrics['wer'] - finetuned_metrics['wer']:+.2f}"
        )
        logger.info(
            f"{'Perfect Matches':<20} {base_metrics['perfect_matches']:<12} {finetuned_metrics['perfect_matches']:<12} {finetuned_metrics['perfect_matches'] - base_metrics['perfect_matches']:+d}"
        )
        logger.info("=" * 80)

    # Create and save model card
    if args.hub_model_id or args.push_to_hub:
        model_id = args.hub_model_id or f"{args.output_dir.split('/')[-1]}"
        logger.info("Creating model card with metrics...")

        model_card_content = create_model_card_content(
            model_id=model_id,
            dataset_id=args.dataset_id,
            subset=args.subset,
            base_metrics=base_metrics,
            finetuned_metrics=finetuned_metrics,
            training_args=training_args,
            freeze_config=freeze_config,
        )

        # Save model card
        model_card_path = Path(args.output_dir) / "README.md"
        model_card_path.write_text(model_card_content)
        logger.info(f"Model card saved to {model_card_path}")

        if args.push_to_hub:
            logger.info(f"Pushing model to Hub: {args.hub_model_id}")
            trainer.push_to_hub()
            logger.info(
                f"✅ Model available at: https://huggingface.co/{args.hub_model_id}"
            )

    logger.info("\n✅ Training complete!")
    logger.info(f"Model saved to: {args.output_dir}")

    # Print example command for inference
    logger.info("\n" + "=" * 80)
    logger.info("To use the fine-tuned model:")
    logger.info("=" * 80)
    logger.info(f"""
from transformers import AutoProcessor, LightOnOCRForConditionalGeneration
from PIL import Image

model = LightOnOCRForConditionalGeneration.from_pretrained("{args.output_dir}")
processor = AutoProcessor.from_pretrained("{args.output_dir}")
# ... rest of inference code
""")


if __name__ == "__main__":
    if len(sys.argv) == 1:
        print("LightOnOCR Fine-tuning Script\n")
        print("Examples:")
        print("  # Basic fine-tuning:")
        print(
            "  uv run lightonocr-finetune.py --dataset-id HuggingFaceM4/FineVision --subset iam --output-dir ./model\n"
        )
        print("  # With frozen components:")
        print(
            "  uv run lightonocr-finetune.py --freeze-language --output-dir ./model\n"
        )
        print("  # Stream large datasets (memory-efficient):")
        print(
            "  uv run lightonocr-finetune.py --streaming --shuffle-buffer-size 10000 --output-dir ./model\n"
        )
        print("  # Push to Hub:")
        print(
            "  uv run lightonocr-finetune.py --hub-model-id username/model --push-to-hub\n"
        )
        print("  # Run on HF Jobs:")
        print(
            "  hf jobs run --gpu l4x1 uv run lightonocr-finetune.py --streaming --output-dir ./model"
        )
        sys.exit(0)

    main()