german-structured-output-olmo2-1b/train.py

"""
German Structured Output SFT Training
======================================
EU AI Act Article 53 compliant fine-tuning of OLMo 2 1B Instruct
on philipp-zettl/german-structured-output dataset.

Base model: allenai/OLMo-2-0425-1B-Instruct (Apache 2.0)
  - Pretraining data: allenai/olmo-mix-1124 (ODC-BY, fully documented)
  - Every data source has explicit license: DCLM (CC-BY-4.0), ArXiv (ODC-BY),
    PeS2o (ODC-BY), StarCoder (ODC-BY), Wikipedia (ODC-BY), etc.
  - Full training logs, code, and intermediate checkpoints published by AI2

Dataset: philipp-zettl/german-structured-output (CC BY-SA 4.0)
  - All sources documented with licenses per example
  - Zero PII, GDPR compliant

Method: Full SFT with gradient checkpointing
"""

import os
import sys
import json
import torch
import statistics

# Force unbuffered output so logs stream in real-time
os.environ["PYTHONUNBUFFERED"] = "1"

print("=" * 60, flush=True)
print("Starting German Structured Output SFT Training", flush=True)
print("=" * 60, flush=True)

# Verify HF_TOKEN is available (needed for private dataset + push_to_hub)
token = os.environ.get("HF_TOKEN")
if not token:
    print("ERROR: HF_TOKEN not set! Cannot access private dataset or push model.", flush=True)
    print("Set via: export HF_TOKEN=hf_... or pass --env HF_TOKEN=hf_...", flush=True)
    sys.exit(1)
else:
    print(f"HF_TOKEN: present ({token[:8]}...)", flush=True)

# Check GPU
print(f"CUDA available: {torch.cuda.is_available()}", flush=True)
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name()}", flush=True)
    print(f"VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB", flush=True)
else:
    print("WARNING: No GPU detected! Training will be very slow.", flush=True)

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import SFTConfig, SFTTrainer

print("All imports OK", flush=True)

# ============================================================================
# Configuration
# ============================================================================

MODEL_ID = "allenai/OLMo-2-0425-1B-Instruct"
DATASET_ID = "philipp-zettl/german-structured-output"
OUTPUT_MODEL_ID = "philipp-zettl/german-structured-output-olmo2-1b"
OUTPUT_DIR = "./output"

# ============================================================================
# Trackio monitoring — writes to HF Bucket for persistent storage
# ============================================================================
print("Initializing Trackio...", flush=True)
import trackio
trackio.init(
    project="german-structured-output-sft",
    name="olmo2-1b-full-sft",
    space_id="philipp-zettl/german-structured-output-training",
    bucket_id="philipp-zettl/german-structured-output-training-bucket",
)
print("Trackio initialized with bucket storage", flush=True)

# ============================================================================
# Load dataset
# ============================================================================
print("Loading dataset...", flush=True)
dataset = load_dataset(DATASET_ID)
print(f"Train: {len(dataset['train'])}, Val: {len(dataset['validation'])}, Test: {len(dataset['test'])}", flush=True)

# Quick data audit
print("\nSample messages structure:", flush=True)
sample = dataset["train"][0]
for msg in sample["messages"]:
    print(f"  [{msg['role']}]: {msg['content'][:100]}...", flush=True)
print(f"  task_type: {sample['task_type']}", flush=True)
print(f"  quality_score: {sample['quality_score']}", flush=True)

# ============================================================================
# Load model and tokenizer
# ============================================================================
print(f"\nLoading model: {MODEL_ID}", flush=True)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa",
)

# Verify chat template exists
print(f"Chat template available: {tokenizer.chat_template is not None}", flush=True)
print(f"Model parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.0f}M", flush=True)

# Check sequence lengths in dataset to set max_length appropriately
print("\nAnalyzing sequence lengths...", flush=True)
lengths = []
for i, example in enumerate(dataset["train"]):
    text = tokenizer.apply_chat_template(example["messages"], tokenize=False)
    tokens = tokenizer(text, return_length=True)
    lengths.append(tokens["length"][0])
    if i >= 200:  # Sample first 200
        break

print(f"  Sampled {len(lengths)} examples", flush=True)
print(f"  Min: {min(lengths)}, Max: {max(lengths)}, Mean: {statistics.mean(lengths):.0f}, Median: {statistics.median(lengths):.0f}", flush=True)
print(f"  P95: {sorted(lengths)[int(len(lengths)*0.95)]}", flush=True)
print(f"  P99: {sorted(lengths)[int(len(lengths)*0.99)]}", flush=True)

# Set max_length to cover P99 + margin, cap at 2048
MAX_LENGTH = min(sorted(lengths)[int(len(lengths)*0.99)] + 128, 2048)
print(f"  Using max_length: {MAX_LENGTH}", flush=True)

# ============================================================================
# Training configuration
# ============================================================================
training_args = SFTConfig(
    output_dir=OUTPUT_DIR,

    # Precision
    bf16=True,

    # Sequence
    max_length=MAX_LENGTH,
    packing=False,  # No packing — variable length structured outputs

    # Batch size: effective = 4 * 4 = 16
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,

    # Optimizer
    learning_rate=2e-5,
    lr_scheduler_type="cosine",
    warmup_ratio=0.05,
    weight_decay=0.01,
    max_grad_norm=1.0,

    # Epochs — small dataset (4K), more epochs to converge
    num_train_epochs=5,

    # Evaluation
    eval_strategy="epoch",
    eval_on_start=True,

    # Saving
    save_strategy="epoch",
    save_total_limit=2,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    greater_is_better=False,

    # Logging — plain text, no tqdm
    logging_strategy="steps",
    logging_steps=10,
    logging_first_step=True,
    disable_tqdm=True,
    report_to="trackio",
    run_name="olmo2-1b-german-structured-output",

    # Memory
    gradient_checkpointing=True,

    # Hub — CRITICAL: push_to_hub to save model (ephemeral job storage)
    push_to_hub=True,
    hub_model_id=OUTPUT_MODEL_ID,
    hub_strategy="end",

    # NEFTune for small dataset regularization (paper: arxiv 2310.05914)
    neftune_noise_alpha=5.0,
)

# ============================================================================
# Trainer
# ============================================================================
print("\nInitializing SFTTrainer...", flush=True)
trainer = SFTTrainer(
    model=model,
    processing_class=tokenizer,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
)

# Print training summary
total_steps = len(dataset["train"]) // (training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps) * int(training_args.num_train_epochs)
print(f"\n{'='*60}", flush=True)
print(f"Training Summary", flush=True)
print(f"{'='*60}", flush=True)
print(f"  Base model: {MODEL_ID}", flush=True)
print(f"  Base model license: Apache 2.0", flush=True)
print(f"  Base model training data: allenai/olmo-mix-1124 (ODC-BY)", flush=True)
print(f"  Dataset: {DATASET_ID}", flush=True)
print(f"  Dataset license: CC BY-SA 4.0", flush=True)
print(f"  Output model: {OUTPUT_MODEL_ID}", flush=True)
print(f"  Train examples: {len(dataset['train'])}", flush=True)
print(f"  Val examples: {len(dataset['validation'])}", flush=True)
print(f"  Max length: {MAX_LENGTH}", flush=True)
print(f"  Effective batch size: {training_args.per_device_train_batch_size} x {training_args.gradient_accumulation_steps} = {training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps}", flush=True)
print(f"  Learning rate: {training_args.learning_rate}", flush=True)
print(f"  Epochs: {training_args.num_train_epochs}", flush=True)
print(f"  Estimated steps: ~{total_steps}", flush=True)
print(f"  NEFTune alpha: {training_args.neftune_noise_alpha}", flush=True)
print(f"  EU AI Act: Article 53 compliant (full data provenance chain)", flush=True)
print(f"{'='*60}", flush=True)

# ============================================================================
# Train!
# ============================================================================
print("\nStarting training...", flush=True)
train_result = trainer.train()

# ============================================================================
# Save and push
# ============================================================================
print("\nSaving model...", flush=True)
trainer.save_model()

# Log final metrics
metrics = train_result.metrics
print(f"\nTraining complete!", flush=True)
print(f"  Train loss: {metrics.get('train_loss', 'N/A')}", flush=True)
print(f"  Train runtime: {metrics.get('train_runtime', 'N/A'):.0f}s", flush=True)
print(f"  Train samples/sec: {metrics.get('train_samples_per_second', 'N/A'):.1f}", flush=True)

# Evaluate
print("\nRunning final evaluation...", flush=True)
eval_metrics = trainer.evaluate()
print(f"  Eval loss: {eval_metrics.get('eval_loss', 'N/A')}", flush=True)

# Push to hub
print(f"\nPushing to Hub: {OUTPUT_MODEL_ID}", flush=True)
trainer.push_to_hub(
    commit_message="EU AI Act compliant SFT: OLMo 2 1B on german-structured-output",
    tags=[
        "german", "structured-output", "json", "function-calling",
        "ner", "relation-extraction", "gdpr-anonymization",
        "eu-ai-act-compliant", "gdpr-compliant", "sft",
        "olmo2", "article-53"
    ],
)

print("\n✅ Training complete and model pushed to Hub!", flush=True)
print(f"   Model: https://huggingface.co/{OUTPUT_MODEL_ID}", flush=True)
print(f"   Trackio: https://huggingface.co/spaces/philipp-zettl/german-structured-output-training", flush=True)