初始化项目，由ModelHub XC社区提供模型

Model: philipp-zettl/german-structured-output-olmo2-1b
Source: Original Platform

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,61 @@
+---
+base_model: allenai/OLMo-2-0425-1B-Instruct
+library_name: transformers
+model_name: german-structured-output-olmo2-1b
+tags:
+- generated_from_trainer
+- trackio
+- hf_jobs
+- trl
+- trackio:https://huggingface.co/spaces/philipp-zettl/huggingface-static-6a4806
+- sft
+licence: license
+---
+
+# Model Card for german-structured-output-olmo2-1b
+
+This model is a fine-tuned version of [allenai/OLMo-2-0425-1B-Instruct](https://huggingface.co/allenai/OLMo-2-0425-1B-Instruct).
+It has been trained using [TRL](https://github.com/huggingface/trl).
+
+## Quick start
+
+```python
+from transformers import pipeline
+
+question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
+generator = pipeline("text-generation", model="philipp-zettl/german-structured-output-olmo2-1b", device="cuda")
+output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
+print(output["generated_text"])
+```
+
+## Training procedure
+
+ 
+
+
+
+This model was trained with SFT.
+
+### Framework versions
+
+- TRL: 1.2.0
+- Transformers: 5.6.1
+- Pytorch: 2.11.0
+- Datasets: 4.8.4
+- Tokenizers: 0.22.2
+
+## Citations
+
+
+
+Cite TRL as:
+    
+```bibtex
+@software{vonwerra2020trl,
+  title   = {{TRL: Transformers Reinforcement Learning}},
+  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
+  license = {Apache-2.0},
+  url     = {https://github.com/huggingface/trl},
+  year    = {2020}
+}
+```

chat_template.jinja

@@ -0,0 +1,9 @@
+{{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %}

config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Olmo2ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 100257,
+  "dtype": "bfloat16",
+  "eos_token_id": 100257,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 4096,
+  "model_type": "olmo2",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "pad_token_id": 100277,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 500000,
+    "rope_type": "default"
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.6.1",
+  "use_cache": false,
+  "vocab_size": 100352
+}

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 100257,
+  "eos_token_id": [
+    100257
+  ],
+  "pad_token_id": 100277,
+  "transformers_version": "5.6.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60e9d083fec2eae48451013015c3c4f0acff6c19acc0e37c03dcd769df25e2ff
+size 2969854224

tokenizer_config.json

@@ -0,0 +1,14 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<|pad|>",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

train.py

@@ -0,0 +1,257 @@
+"""
+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)

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:022802dfe5deb48d1b90fab3e0b01f9e32604a91ae639f58a6adc24272409a80
+size 5841