| |
| """Publish verified local artifacts without provisioning any inference service.""" |
| from __future__ import annotations |
|
|
| import argparse |
| import getpass |
| import hashlib |
| import json |
| import math |
| import os |
| import re |
| import sys |
| import tempfile |
| from datetime import datetime, timezone |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| sys.path.insert(0, str(ROOT / 'scripts')) |
|
|
|
|
| def sha256(path): |
| h = hashlib.sha256() |
| with Path(path).open('rb') as f: |
| for block in iter(lambda: f.read(8 * 1024 * 1024), b''): |
| h.update(block) |
| return h.hexdigest() |
|
|
|
|
| def read_json(path): |
| return json.loads(Path(path).read_text()) |
|
|
|
|
| def read_jsonl(path): |
| return [json.loads(x) for x in Path(path).read_text().splitlines() if x.strip()] |
|
|
|
|
| def safe_member(root, name): |
| target = (root / name).resolve() |
| if not target.is_relative_to(root.resolve()) or not target.is_file(): |
| raise ValueError(f'Missing or unsafe artifact file: {name}') |
| return target |
|
|
|
|
| def license_files(directory): |
| """Require the full Apache license and preserve every existing notice verbatim.""" |
| directory = Path(directory) |
| main = safe_member(directory, 'LICENSE') |
| text = main.read_text(encoding='utf-8') |
| if (len(text) < 9000 or 'Apache License' not in text or 'Version 2.0, January 2004' not in text |
| or 'END OF TERMS AND CONDITIONS' not in text |
| or any(not re.search(rf'(?m)^\s*{number}\.\s+\S', text) for number in range(1, 10))): |
| raise ValueError('A full original Apache-2.0 LICENSE is required before upload') |
| return {path.name: safe_member(directory, path.name) for path in sorted(directory.iterdir()) |
| if path.is_file() and path.name.startswith(('LICENSE', 'NOTICE'))} |
|
|
|
|
| def prepare_gguf_licenses(merged, gguf): |
| """Copy only missing upstream notices; never replace a conflicting file.""" |
| originals = license_files(merged) |
| existing = {path.name for path in gguf.iterdir() if path.is_file() and path.name.startswith(('LICENSE', 'NOTICE'))} |
| if existing - set(originals): |
| raise ValueError('GGUF contains license or notice files absent from the upstream export') |
| for name, original in originals.items(): |
| target = gguf / name |
| if target.exists() and (not target.is_file() or target.read_bytes() != original.read_bytes()): |
| raise ValueError('GGUF license or notice differs from the original upstream file') |
| for name, original in originals.items(): |
| target = gguf / name |
| if not target.exists(): |
| with target.open('xb') as handle: |
| handle.write(original.read_bytes()) |
| handle.flush() |
| os.fsync(handle.fileno()) |
| return originals |
|
|
|
|
| def upload_licenses(api, repo, repo_type, files): |
| for name, path in files.items(): |
| api.upload_file(repo_id=repo, repo_type=repo_type, path_or_fileobj=str(path), path_in_repo=name, |
| commit_message='Preserve complete original license and notice files') |
|
|
|
|
| def inspect_evaluation(summary_path, data, adapter): |
| summary = read_json(summary_path) |
| if summary.get('status') != 'complete' or summary.get('expected_prompts') != 500 or summary.get('recorded_prompts') != 500 or summary.get('infrastructure_failures', 1) != 0: |
| raise ValueError('Publication requires a complete 500-prompt evaluation without infrastructure failures') |
| if summary.get('run_seconds') != 60 or summary.get('pass_k') != 1 or summary.get('retries') != 0: |
| raise ValueError('Publication requires unrepaired pass@1 with 60-second runtime windows') |
| evaluation = summary_path.parent |
| manifest = read_json(evaluation / 'manifest.json') |
| if manifest.get('smoke_test', True) or manifest.get('expected_count') != 500 or manifest.get('run_seconds') != 60: |
| raise ValueError('A smoke-test evaluation cannot be used for publication') |
| if manifest.get('sample_indices') != [0, 200, 400] or summary.get('sample_indices') != [0, 200, 400]: |
| raise ValueError('Evaluation does not use the predeclared sample indices 0, 200, and 400') |
| if manifest.get('prompts_sha256') != sha256(data / 'eval_prompts.jsonl'): |
| raise ValueError('Evaluation prompt hash differs from frozen dataset') |
| prompts = read_jsonl(data / 'eval_prompts.jsonl') |
| if len(prompts) != 500 or len({row['instruction'] for row in prompts}) != 500: |
| raise ValueError('Dataset does not contain 500 unique evaluation prompts') |
| |
| expected_adapter = manifest.get('generation', {}).get('adapter', {}).get('files', {}) |
| if expected_adapter.get('adapter_model.safetensors') != sha256(adapter / 'adapter_model.safetensors'): |
| raise ValueError('Evaluated adapter does not match the adapter being uploaded') |
| files = sorted((evaluation / 'items').glob('*/result.json')) |
| if len(files) != 500: |
| raise ValueError('Evaluation summary is missing underlying per-prompt evidence') |
| results = [read_json(p) for p in files] |
| if sorted(r.get('index', -1) for r in results) != list(range(500)): |
| raise ValueError('Evaluation indices are incomplete or duplicated') |
| compiled, clean = 0, 0 |
| for path, result in zip(files, results): |
| attempt = read_json(path.parent / 'attempt.json') |
| index = result['index'] |
| if attempt.get('index') != index or attempt.get('instruction') != prompts[index]['instruction']: |
| raise ValueError('Per-prompt evidence does not match the frozen prompt ordering') |
| if result.get('infrastructure_failure'): |
| raise ValueError('Infrastructure failures cannot be reported as model metrics') |
| declared = result.get('validation_evidence') |
| |
| |
| completed_check = None |
| if declared is not None: |
| if not isinstance(declared, str) or not re.fullmatch(r'check(?:-resume-[0-9]{3,})?', declared): |
| raise ValueError('Unsafe validation evidence directory') |
| completed_check = (path.parent / declared).resolve() |
| if not completed_check.is_relative_to(path.parent.resolve()) or not (completed_check / 'validation.json').is_file(): |
| raise ValueError('Validation evidence pointer is missing or leaves the evaluated item') |
| checks = [x.resolve().parent for x in path.parent.glob('check*/validation.json') if x.is_file()] |
| if completed_check is not None: |
| if len(checks) != 1 or checks[0] != completed_check: |
| raise ValueError('Validation evidence pointer does not match the sole completed attempt') |
| elif len(checks) == 1 and checks[0].is_relative_to(path.parent.resolve()): |
| completed_check = checks[0] |
| if result.get('clean_run_60s') and not result.get('compile_success'): |
| raise ValueError('A clean runtime cannot be reported after failed compilation') |
| if result.get('compile_success'): |
| compiled += 1 |
| if completed_check is None or not (completed_check / 'import.log').exists(): |
| raise ValueError('Successful compile lacks unambiguous import log evidence') |
| if result.get('clean_run_60s'): |
| clean += 1 |
| check = read_json(completed_check / 'validation.json') |
| if not check.get('compile_success') or not check.get('clean_run_60s'): |
| raise ValueError('Clean-run result contradicts validation evidence') |
| runtime = check.get('runtime', {}) |
| if runtime.get('wall_seconds', 0) < 60: |
| raise ValueError('Successful runtime was shorter than 60 seconds') |
| runtime_log = completed_check / 'runtime.log' |
| if not runtime_log.exists() or runtime.get('log_sha256') != sha256(runtime_log): |
| raise ValueError('Clean runtime lacks matching log evidence') |
| for name, count in [('compile_success', compiled), ('clean_run_60s', clean)]: |
| if summary.get(name + '_count') != count or not math.isclose(summary.get(name + '_rate', -1), count / 500, abs_tol=1e-12): |
| raise ValueError('Reported metrics disagree with per-prompt evidence') |
| return summary, manifest |
|
|
|
|
| def inspect_weights(adapter, merged, gguf): |
| license_files(merged) |
| from safetensors import safe_open |
| training = read_json(adapter / 'TRAINING_COMPLETE.json') |
| if training.get('status') != 'complete' or training.get('global_step', 0) <= 0 or training.get('global_step') != training.get('target_steps'): |
| raise ValueError('Adapter training is incomplete') |
| config = read_json(adapter / 'adapter_config.json') |
| if config.get('r') != 32 or config.get('lora_alpha') != 64: |
| raise ValueError('Unexpected LoRA rank or alpha') |
| adapter_file = adapter / 'adapter_model.safetensors' |
| with safe_open(adapter_file, framework='pt', device='cpu') as f: |
| keys = list(f.keys()) |
| adapter_parameters = sum(math.prod(f.get_slice(k).get_shape()) for k in keys) |
| if len(keys) != 392 or adapter_parameters < 1_000_000: |
| raise ValueError('Adapter does not contain the expected 28-layer, seven-module rank-32 tensors') |
| merged_marker = read_json(merged / 'MERGE_COMPLETE.json') |
| if merged_marker.get('adapter_sha256') != sha256(adapter_file): |
| raise ValueError('Merged model was produced from a different adapter') |
| index = read_json(merged / 'model.safetensors.index.json')['weight_map'] |
| parameters = 0 |
| by_file = {} |
| for name, filename in index.items(): |
| by_file.setdefault(filename, []).append(name) |
| for filename, names in by_file.items(): |
| with safe_open(safe_member(merged, filename), framework='pt', device='cpu') as f: |
| if set(f.keys()) != set(names): |
| raise ValueError('Merged tensor index and shard contents disagree') |
| for name in names: |
| tensor = f.get_slice(name) |
| if tensor.get_dtype() != 'F16': |
| raise ValueError('Merged model contains a non-fp16 tensor') |
| parameters += math.prod(tensor.get_shape()) |
| if not 7_000_000_000 < parameters < 8_000_000_000: |
| raise ValueError('Merged artifact is not a complete Qwen2.5-Coder-7B-size model') |
| gguf_file = gguf / 'VibeGame-8B-Q4_K_M.gguf' |
| with gguf_file.open('rb') as f: |
| if f.read(4) != b'GGUF': |
| raise ValueError('GGUF file has invalid magic bytes') |
| quant = read_json(gguf / 'GGUF_COMPLETE.json') |
| if quant.get('quantization') != 'Q4_K_M' or quant.get('sha256') != sha256(gguf_file) or gguf_file.stat().st_size < 1_000_000_000: |
| raise ValueError('GGUF is incomplete or does not match its export manifest') |
| if quant.get('source_merge', {}).get('adapter_sha256') != sha256(adapter_file): |
| raise ValueError('GGUF export was produced from a different adapter') |
| return {'training': training, 'adapter_sha256': sha256(adapter_file), |
| 'merged_parameters': parameters, 'gguf_sha256': quant['sha256']} |
|
|
|
|
| def inspect_artifacts(adapter, merged, gguf, data, summary_path): |
| license_files(ROOT) |
| weights = inspect_weights(adapter, merged, gguf) |
| dataset = read_json(data / 'dataset_manifest.json') |
| rows = read_jsonl(data / 'train.jsonl') + read_jsonl(data / 'dev.jsonl') |
| if dataset.get('train', 0) + dataset.get('dev', 0) != len(rows) or not rows: |
| raise ValueError('Dataset counts are inconsistent') |
| admitted = {r['output_sha256'] for r in read_jsonl(data / 'validation_manifest.jsonl') if r.get('accepted')} |
| if any(hashlib.sha256(r['output'].encode()).hexdigest() not in admitted for r in rows): |
| raise ValueError('Dataset contains a project without accepted Godot validation') |
| for row in read_jsonl(data / 'reference_gd.jsonl'): |
| if row.get('license') not in {'MIT', 'Apache-2.0', 'CC0-1.0'}: |
| raise ValueError('Dataset reference corpus contains a disallowed license') |
| if not (data / 'licenses').is_dir() or not list((data / 'licenses').rglob('*')): |
| raise ValueError('Source license notices are missing') |
| summary, evaluation_manifest = inspect_evaluation(summary_path, data, adapter) |
| return {**weights, 'dataset': dataset, 'evaluation': summary, |
| 'evaluation_manifest': evaluation_manifest} |
|
|
|
|
| def provider_mapping(api, repo): |
| info = api.model_info(repo_id=repo, expand=['inferenceProviderMapping']) |
| mapping = getattr(info, 'inference_provider_mapping', None) |
| if mapping is None: |
| mapping = getattr(info, 'inferenceProviderMapping', None) |
| |
| if mapping is None: |
| raise RuntimeError('Hugging Face did not return inferenceProviderMapping; cannot verify provider state') |
| if not isinstance(mapping, (dict, list)): |
| raise TypeError('Unexpected inference provider response') |
| return mapping |
|
|
|
|
| def model_card(repo_ids, evidence, template=None): |
| e, d = evidence['evaluation'], evidence['dataset'] |
| gpu = evidence.get('evaluation_manifest', {}).get('generation', {}).get('gpu', 'Not recorded') |
| hardware = re.sub(r'[\r\n`|]', ' ', str(gpu))[:120] |
| context = ('This table reports the separate full A100 pass on Colab Pro. The partial T4 experiment is ' |
| 'preserved under `evaluation-t4-partial/`; its results are not pooled with this pass.' |
| if 'A100' in hardware else 'The hardware and runtime versions for this pass are recorded in `evaluation/manifest.json`.') |
| demo_selection = ('This video is a success-only showcase: exactly the first three clean results in frozen prompt order. ' |
| 'It is not representative of the model. All three recordings use untouched outputs, and the full ' |
| '500-prompt results and actual times stay visible. The original benchmark sample slots 0, 200, ' |
| 'and 400 remain unchanged in the evidence.' |
| if evidence.get('demo_kind') == 'model-success-showcase' else |
| 'Samples 0, 200, and 400 were chosen before evaluation. Failed samples keep their slots as diagnostic cards.') |
| values = {'COMPILE_RATE': f"{e['compile_success_rate']:.1%}", 'CLEAN_RUN_RATE': f"{e['clean_run_60s_rate']:.1%}", |
| 'COMPILE_COUNT': str(e['compile_success_count']), 'CLEAN_RUN_COUNT': str(e['clean_run_60s_count']), |
| 'UNDER_60_COUNT': str(e['under_60s_launch_count']), |
| 'TRAIN_COUNT': str(d['train']), 'DEV_COUNT': str(d['dev']), |
| 'MEDIAN_SECONDS': str(e['prompt_to_launch_median_seconds']), |
| 'MEDIAN_LATENCY': (f"{e['prompt_to_launch_median_seconds']:.2f} seconds" |
| if e['prompt_to_launch_median_seconds'] is not None else 'Not measured'), |
| 'EVAL_HARDWARE': hardware, 'EVAL_CONTEXT': context, |
| 'DEMO_SELECTION': demo_selection, |
| 'MODEL_REPO': repo_ids['merged'], 'GGUF_REPO': repo_ids['gguf'], 'ADAPTER_REPO': repo_ids['adapter'], 'DATASET_REPO': repo_ids['dataset']} |
| if template and template.exists(): |
| body = template.read_text() |
| |
| if body.startswith('---\n') and '\n---\n' in body[4:]: |
| body = body.split('\n---\n', 1)[1] |
| for key, value in values.items(): |
| body = body.replace('{{' + key + '}}', value) |
| if re.search(r'\{\{[A-Z0-9_]+\}\}', body): |
| raise ValueError('Model card contains unresolved release placeholders') |
| else: |
| body = f'''# VibeGame-8B |
| |
| I wanted a small open model that could turn a game idea into a Godot project I could actually open and play. This is the experiment: describe a 2D game and get project.godot, a scene, and GDScript in one JSON response. |
| |
| The project name is VibeGame-8B. The base is Qwen2.5-Coder-7B-Instruct, with 7.61 billion parameters. “Prompt to playable game in 60 seconds” is the target. The measured timing below says how close this release gets; it is not a blanket promise. |
| |
| Download the [GGUF](https://huggingface.co/{repo_ids['gguf']}) and its Modelfile. Run `ollama create vibegame-8b -f Modelfile`, then use the repository generator and validator. Generated code should pass the sandboxed Godot check before you open it. This model uses Godot 4.5.1 and procedural shapes, with no game assets bundled. |
| |
| The [dataset](https://huggingface.co/datasets/{repo_ids['dataset']}) has {d['train']} training pairs and {d['dev']} development pairs. Whole synthetic projects must import and run cleanly for 60 seconds before admission. The reference corpus accepts MIT, Apache-2.0, and CC0 code, retains source notices, and excludes game assets. Deduplication happens before splitting. The 500 frozen evaluation prompts never enter the training loss. |
| |
| These are small games. A script can run cleanly while the controls, scoring, or win condition are wrong. The benchmark checks parsing and runtime behavior without pressing buttons; it does not replace playtesting. Longer projects may exceed the 4096-token context, and JSON can be incomplete. We do not repair outputs during evaluation or hide failures. |
| |
| If you want to help, send a small reproducible failing prompt, a permissively licensed project with clear provenance, or a validator improvement. More independently authored game mechanics and actual playtesting would help more than another favorable screenshot. |
| ''' |
| body += f''' |
| |
| ## Measured release results |
| |
| | Check | Result | |
| | --- | --- | |
| | Frozen prompts | 500 | |
| | Evaluation GPU | {hardware} | |
| | Compile success | {e['compile_success_count']}/500 ({e['compile_success_rate']:.1%}) | |
| | Clean 60-second run | {e['clean_run_60s_count']}/500 ({e['clean_run_60s_rate']:.1%}) | |
| | Median prompt-to-launch among compiled games | {values['MEDIAN_LATENCY']} | |
| | Compiled games launched within 60 seconds | {e['under_60s_launch_count']}/500 | |
| | Sampling | One generation per prompt; no repairs or retries | |
| |
| These numbers evaluate the **adapter on its 4-bit base**. They are not measurements of the merged fp16 or GGUF artifacts. Export changes numerical precision; those formats still need a separate run. Three sample indices are fixed at 0, 200, and 400 before evaluation, covering distinct game families. See `evaluation/summary.json` and the full per-prompt evidence for every failure. |
| |
| {context} |
| |
| {demo_selection} |
| |
| The provider mapping was checked during publication. `inference: false` hides the widget; it is not an account billing switch. No endpoint or paid inference service is created by this release script. Account provider/billing settings are verified separately. |
| ''' |
| header = '---\nlibrary_name: transformers\nlicense: apache-2.0\nbase_model: Qwen/Qwen2.5-Coder-7B-Instruct\npipeline_tag: text-generation\ninference: false\ntags:\n- godot\n- gdscript\n- qlora\n- game-development\n---\n\n' |
| return header + body.rstrip() + '\n' |
|
|
|
|
| def sanitized_text(text): |
| |
| |
| text = re.sub(r"/Users/[^/\\\"\s]+", "/redacted-home", text) |
| text = re.sub(r"/home/[^/\\\"\s]+", "/redacted-home", text) |
| return text.replace('/content/drive/MyDrive', '/persistent-drive') |
|
|
|
|
| def public_evidence_text(text): |
| """Reject recognizable credentials instead of uploading or printing them.""" |
| if re.search(r'\bhf_[A-Za-z0-9]{20,}\b|\bsk-[A-Za-z0-9_-]{20,}\b|\bBearer\s+[A-Za-z0-9._~-]{12,}', text, re.IGNORECASE): |
| raise ValueError('Publication evidence contains a possible credential; nothing was uploaded') |
| try: |
| value = json.loads(text) |
| except ValueError: |
| value = None |
| sensitive = {'token', 'hf_token', 'access_token', 'authorization', 'password', 'api_key', 'secret', 'client_secret'} |
| def check(item): |
| if isinstance(item, dict): |
| for key, child in item.items(): |
| if str(key).lower() in sensitive and child not in (None, '', False): |
| raise ValueError('Publication evidence contains a credential field; nothing was uploaded') |
| check(child) |
| elif isinstance(item, list): |
| for child in item: |
| check(child) |
| check(value) |
| return sanitized_text(text) |
|
|
|
|
| def a100_receipts(evaluation, manifest, data, adapter): |
| """Bind the extra A100 receipts to this pass and its preserved T4 source.""" |
| if evaluation.name != 'evaluation-a100-v1' or evaluation.is_symlink(): |
| raise ValueError('A100 release requires its separate evaluation-a100-v1 directory') |
| declaration_path = safe_member(evaluation, 'pro-protocol.json') |
| declaration = read_json(declaration_path) |
| protocol = declaration['protocol'] |
| generation = manifest['generation'] |
| prior = evaluation.parent / 'evaluation' |
| if prior.is_symlink(): |
| raise ValueError('T4 evidence directory must not be a symlink') |
| prior_manifest_path = safe_member(prior, 'manifest.json') |
| prior_manifest = read_json(prior_manifest_path) |
| if (protocol.get('experiment') != 'evaluation-a100-v1' |
| or protocol.get('source_t4_manifest_sha256') != sha256(prior_manifest_path) |
| or 'T4' not in prior_manifest.get('generation', {}).get('gpu', '') |
| or 'A100' not in generation.get('gpu', '') |
| or protocol.get('gpu_runtime', {}).get('gpu') != generation['gpu'] |
| or protocol.get('train_sha256') != sha256(data / 'train.jsonl')): |
| raise ValueError('A100 predeclaration does not match the selected experiment or T4 source') |
| for key in ('expected_count', 'batch_size', 'workers', 'max_new_tokens', 'run_seconds', 'sample_indices', 'decoding', 'prompts_sha256'): |
| if protocol.get(key) != manifest.get(key) or manifest.get(key) != prior_manifest.get(key): |
| raise ValueError('A100 and T4 protocols differ: ' + key) |
| if protocol.get('godot_sha256') != manifest.get('runtime_sha256') or manifest.get('runtime_sha256') != prior_manifest.get('runtime_sha256'): |
| raise ValueError('A100 and T4 Godot fingerprints differ') |
| for key in ('adapter', 'base_model', 'base_revision', 'system_prompt_sha256'): |
| if protocol.get(key) != generation.get(key) or generation.get(key) != prior_manifest['generation'].get(key): |
| raise ValueError('A100 and T4 model identities differ: ' + key) |
| for key in ('torch_version', 'cuda_version', 'packages'): |
| if protocol.get('gpu_runtime', {}).get(key) != generation.get(key) or generation.get(key) != prior_manifest['generation'].get(key): |
| raise ValueError('A100 and T4 runtime versions differ: ' + key) |
| declared = datetime.fromisoformat(declaration['declared_utc']) |
| if declared.tzinfo is None: |
| raise ValueError('A100 predeclaration needs an explicit timezone') |
| first_attempt = min(datetime.fromisoformat(read_json(path)['started_utc']) for path in evaluation.glob('items/*/attempt.json')) |
| if declared > first_attempt: |
| raise ValueError('A100 protocol was declared after blind evaluation started') |
| |
| from benchmark_batch import capacity_prompts |
| from evaluate_pro import capacity_valid |
| training_prompts = [row['instruction'] for row in capacity_prompts(adapter, data / 'train.jsonl')] |
| capacities, receipts = {}, [declaration_path] |
| for pattern in ('capacity-*.json', 'budget-*.json'): |
| for path in sorted(evaluation.glob(pattern)): |
| if not re.fullmatch(r'(?:capacity|budget)-[0-9]{4,}\.json', path.name): |
| raise ValueError('Unexpected A100 receipt filename') |
| path = safe_member(evaluation, path.name) |
| public_evidence_text(path.read_text()) |
| receipt = read_json(path) |
| receipts.append(path) |
| if path.name.startswith('capacity-'): |
| capacities[sha256(path)] = receipt |
| budgets = [read_json(path) for path in receipts if path.name.startswith('budget-')] |
| approved = [] |
| for budget in budgets: |
| if budget.get('fits_observed_balance') is True: |
| capacity = capacities.get(budget.get('capacity_report_sha256')) |
| if capacity is None or not capacity_valid(capacity, protocol, prior_manifest['generation'], training_prompts): |
| raise ValueError('Approved A100 budget has no matching valid capacity report') |
| approved.append(budget) |
| if not any(datetime.fromisoformat(row['observed_utc']) <= first_attempt for row in approved): |
| raise ValueError('A100 release lacks an approved pre-evaluation capacity/budget receipt') |
| return receipts, prior |
|
|
|
|
| def evaluation_uploads(summary_path, data, adapter): |
| """Select and screen public evidence before requesting a publication token.""" |
| evaluation = summary_path.parent |
| manifest = read_json(evaluation / 'manifest.json') |
| sources = [('evaluation', evaluation, [])] |
| if ('A100' in manifest.get('generation', {}).get('gpu', '') |
| or evaluation.name == 'evaluation-a100-v1' or (evaluation / 'pro-protocol.json').exists()): |
| receipts, prior = a100_receipts(evaluation, manifest, data, adapter) |
| sources = [('evaluation', evaluation, receipts), ('evaluation-t4-partial', prior, [])] |
| allowed_names = {'summary.json', 'manifest.json', 'result.json', 'attempt.json', 'generation.json', |
| 'validation.json', 'interruption.json', 'import.log', 'runtime.log'} |
| uploads = [] |
| for prefix, directory, extra in sources: |
| extra = {path.resolve() for path in extra} |
| files = {} |
| for local in sorted(directory.rglob('*')): |
| if not local.is_file(): |
| continue |
| relative = local.relative_to(directory) |
| showcase_file = (str(relative) == 'showcase/selection.json' |
| or (relative.parts[:2] == ('showcase', 'recordings') |
| and re.fullmatch(r'sample-[0-9]{3}\.mp4|sample-[0-9]{3}\.recording\.json', local.name))) |
| if not (local.name in allowed_names or local.resolve() in extra |
| or showcase_file or (relative.parts[0] == 'recordings' and local.suffix in {'.mp4', '.json'})): |
| continue |
| if local.is_symlink() or not local.resolve().is_relative_to(directory.resolve()): |
| raise ValueError('Unsafe public evaluation evidence path') |
| if local.suffix != '.mp4': |
| public_evidence_text(local.read_text()) |
| files[str(relative)] = {'path': local, 'sha256': sha256(local)} |
| if not {'summary.json', 'manifest.json'} <= files.keys(): |
| raise ValueError('Selected evaluation evidence lacks its manifest or summary') |
| uploads.append({'prefix': prefix, 'files': files}) |
| return uploads |
|
|
|
|
| def stage_evaluation_upload(upload, destination): |
| redactions = {} |
| for relative, record in upload['files'].items(): |
| local = record['path'] |
| if sha256(local) != record['sha256']: |
| raise ValueError('Evaluation evidence changed after publication preflight') |
| target = destination / relative |
| target.parent.mkdir(parents=True, exist_ok=True) |
| if local.suffix == '.mp4': |
| target.symlink_to(local.resolve()) |
| else: |
| original = local.read_text() |
| redacted = public_evidence_text(original) |
| target.write_text(redacted) |
| if original != redacted: |
| redactions[relative] = {'original_sha256': record['sha256'], 'published_sha256': sha256(target), |
| 'change': 'Local home-path prefix redaction only; original evidence hashes retained in manifests.'} |
| (destination / 'REDACTIONS.json').write_text(json.dumps(redactions, indent=2) + '\n') |
|
|
|
|
| def inspect_demo(summary_path): |
| """Require the launch asset produced from this evaluation and fixed samples.""" |
| video = summary_path.parent.parent / 'VibeGame-launch-60s.mp4' |
| sidecar = video.with_suffix('.demo.json') |
| if not video.is_file() or not sidecar.is_file(): |
| raise ValueError('Publication requires the launch video and its demo evidence') |
| edit = read_json(sidecar) |
| if edit.get('scaffold_preview') is not False or edit.get('output_sha256') != sha256(video) or edit.get('total_seconds') != 60: |
| raise ValueError('Launch video lacks matching final-model evidence') |
| plan = edit.get('source_plan', {}) |
| if edit.get('kind') == 'model-success-showcase': |
| from build_demo import evaluation_plan |
| from showcase import selection_path |
| expected = evaluation_plan(summary_path.parent, showcase_selection=selection_path(summary_path.parent)) |
| if (plan != expected or edit.get('watermark') != expected['selection_disclosure'] |
| or len(plan['samples']) != 3 or not all(sample.get('capture') for sample in plan['samples'])): |
| raise ValueError('Showcase must match the exact first three clean results and all three original captures') |
| return video, sidecar |
| if (edit.get('kind') != 'model-evaluation' or plan.get('kind') != 'model-evaluation' |
| or plan.get('scaffold_preview') is not False |
| or plan.get('evaluation_manifest_sha256') != sha256(summary_path.parent / 'manifest.json')): |
| raise ValueError('Launch video belongs to a different evaluation or is a scaffold preview') |
| samples = plan.get('samples', []) |
| if not isinstance(samples, list) or [sample.get('index') for sample in samples if isinstance(sample, dict)] != [0, 200, 400] or len(samples) != 3: |
| raise ValueError('Launch video does not use predeclared samples 0, 200, and 400') |
| if not any(sample.get('capture') for sample in samples): |
| raise ValueError('Launch video requires at least one actual gameplay capture') |
| return video, sidecar |
|
|
|
|
| def main(): |
| p = argparse.ArgumentParser(description=__doc__) |
| for name in ('adapter', 'merged', 'gguf', 'data', 'evaluation', 'output'): |
| p.add_argument('--' + name, type=Path, required=True) |
| p.add_argument('--namespace', required=True) |
| p.add_argument('--model-card', type=Path, default=ROOT / 'docs/model_card.md') |
| args = p.parse_args() |
| if not re.fullmatch(r'[A-Za-z0-9][A-Za-z0-9_-]*', args.namespace): |
| p.error('Invalid Hugging Face namespace') |
| |
| evidence = inspect_artifacts(args.adapter, args.merged, args.gguf, args.data, args.evaluation) |
| upstream_licenses = prepare_gguf_licenses(args.merged, args.gguf) |
| project_licenses = license_files(ROOT) |
| video, video_evidence = inspect_demo(args.evaluation) |
| evidence['demo_kind'] = read_json(video_evidence).get('kind') |
| uploads = evaluation_uploads(args.evaluation, args.data, args.adapter) |
| repo_ids = {'merged': f'{args.namespace}/VibeGame-8B', 'gguf': f'{args.namespace}/VibeGame-8B-GGUF', |
| 'adapter': f'{args.namespace}/VibeGame-8B-LoRA', 'dataset': f'{args.namespace}/VibeGame-8B-data'} |
| card = model_card(repo_ids, evidence, args.model_card) |
| from stage_exports import PublicationTracker, atomic_json, build_manifests |
| previous_report = read_json(args.output) if args.output.exists() else {} |
| from huggingface_hub import HfApi |
| token = os.environ.get('HF_TOKEN') or getpass.getpass('Hugging Face write token (not saved): ') |
| if not token: |
| raise RuntimeError('A write token is required to publish; nothing was uploaded') |
| api = HfApi(token=token) |
| release_identity = hashlib.sha256(json.dumps({ |
| 'namespace': args.namespace, 'adapter_sha256': evidence['adapter_sha256'], |
| 'adapter_config_sha256': sha256(args.adapter / 'adapter_config.json'), |
| 'gguf_sha256': evidence['gguf_sha256'], 'merge_receipt_sha256': sha256(args.merged / 'MERGE_COMPLETE.json'), |
| 'evaluation_manifest_sha256': sha256(args.evaluation.parent / 'manifest.json'), |
| 'upstream_licenses': {name: sha256(path) for name, path in upstream_licenses.items()}, |
| 'project_licenses': {name: sha256(path) for name, path in project_licenses.items()}, |
| 'data_sha256': {name: sha256(args.data / name) for name in ('train.jsonl', 'dev.jsonl', 'eval_prompts.jsonl')}, |
| }, sort_keys=True).encode()).hexdigest() |
| report = {'status': 'uploading', 'repos': repo_ids, 'created_utc': datetime.now(timezone.utc).isoformat(), |
| 'release_identity': release_identity, |
| 'private_staging': dict(previous_report.get('private_staging', {})), |
| 'publication_ownership': dict(previous_report.get('publication_ownership', {})), |
| 'artifacts_verified': True, 'provider_checks': {}, 'paid_inference_calls': 0, |
| 'endpoint_creation_calls': 0, 'billing_changes': 0, |
| 'account_billing_status': 'Requires post-publication account settings verification; provider mappings alone do not prove account billing disabled.'} |
| report['evaluation_evidence'] = {upload['prefix']: {relative: record['sha256'] for relative, record in upload['files'].items()} |
| for upload in uploads} |
| args.output.parent.mkdir(parents=True, exist_ok=True) |
| def save_report(): |
| |
| |
| atomic_json(args.output, json.loads(json.dumps(report, default=str))) |
| api = PublicationTracker(api, report, save_report) |
| try: |
| identity = api.whoami() |
| allowed = {identity['name']} | {x['name'] for x in identity.get('orgs', [])} |
| if args.namespace not in allowed: |
| raise RuntimeError('Token identity has no membership in the requested namespace') |
| existing = {} |
| manifests = None |
| for kind, repo in repo_ids.items(): |
| repo_type = 'dataset' if kind == 'dataset' else 'model' |
| exists = api.repo_exists(repo_id=repo, repo_type=repo_type) |
| existing[kind] = exists |
| if exists: |
| info = api.repo_info(repo_id=repo, repo_type=repo_type) |
| if info.private is True: |
| if kind not in {'merged', 'gguf'}: |
| raise ValueError('Refusing to publish an unrelated private repository') |
| if manifests is None: |
| manifests = build_manifests(args.adapter, args.merged, args.gguf, args.namespace, weights=evidence) |
| api.adopt(manifests[kind], previous_report.get('private_staging', {}).get(repo)) |
| elif info.private is False: |
| previous = (previous_report.get('publication_ownership', {}).get(repo) |
| or previous_report.get('private_staging', {}).get(repo)) |
| manifest = None |
| if previous and previous.get('manifest_sha256'): |
| if kind not in {'merged', 'gguf'}: |
| raise ValueError('Staging identity is invalid for this public repository') |
| if manifests is None: |
| manifests = build_manifests(args.adapter, args.merged, args.gguf, args.namespace, weights=evidence) |
| manifest = manifests[kind] |
| api.adopt_public(repo, repo_type, previous, manifest) |
| else: |
| raise ValueError('Repository visibility could not be verified') |
| for kind, repo in repo_ids.items(): |
| repo_type = 'dataset' if kind == 'dataset' else 'model' |
| if not existing[kind]: |
| api.create_repo(repo_id=repo, repo_type=repo_type, private=False, exist_ok=False) |
| api.track_created(repo, repo_type) |
| if repo_type == 'model': |
| mapping = provider_mapping(api, repo) |
| report['provider_checks'][repo] = {'before': mapping} |
| if mapping: |
| raise RuntimeError(f'Provider mapping already active for {repo}; refusing upload') |
| save_report() |
| folders = {'merged': args.merged, 'gguf': args.gguf, 'adapter': args.adapter, 'dataset': args.data} |
| patterns = {'merged': ['*.json', '*.safetensors', '*.txt', '*.model', '*.jinja', 'LICENSE*', 'NOTICE*'], |
| 'gguf': ['*.gguf', '*.json', 'Modelfile', 'LICENSE*', 'NOTICE*'], |
| 'adapter': ['*.json', '*.safetensors', '*.txt', '*.model', '*.jinja'], |
| 'dataset': ['*.jsonl', '*.json', 'licenses/**']} |
| for kind, repo in repo_ids.items(): |
| repo_type = 'dataset' if kind == 'dataset' else 'model' |
| api.upload_folder(repo_id=repo, repo_type=repo_type, folder_path=folders[kind], |
| allow_patterns=patterns[kind], ignore_patterns=['**/.cache/**', '*.tmp', 'merge_progress.json', '*manifest.json', 'teacher_provenance.json', 'teacher_attempts.jsonl', 'validation_manifest.jsonl', 'rejected_candidates.jsonl'], |
| commit_message='Publish validated VibeGame research artifacts') |
| |
| |
| upload_licenses(api, repo, repo_type, project_licenses if kind == 'dataset' else upstream_licenses) |
| for local in folders[kind].glob('*'): |
| if local.is_file() and (local.name.endswith('manifest.json') or local.name in {'teacher_provenance.json', 'teacher_attempts.jsonl', 'validation_manifest.jsonl', 'rejected_candidates.jsonl'}): |
| api.upload_file(repo_id=repo, repo_type=repo_type, path_or_fileobj=sanitized_text(local.read_text()).encode(), |
| path_in_repo=local.name, commit_message='Add provenance with local home paths removed') |
| if kind == 'dataset': |
| dataset_card = '---\nlicense: apache-2.0\ntask_categories:\n- text-generation\nlanguage:\n- en\n---\n\n# VibeGame dataset\n\nOriginal synthetic projects plus a separately stored permissive GDScript reference corpus. Source files keep their original MIT, Apache-2.0, or CC0 terms and notices in `licenses/`; this repository does not relicense them. `train.jsonl` and `dev.jsonl` contain only validated projects. `eval_prompts.jsonl` is a frozen 500-prompt benchmark, never used for loss. See provenance and validation manifests for accepted and rejected examples. No game assets are included.\n' |
| content = dataset_card |
| else: |
| content = card |
| if kind == 'adapter': |
| content = content.replace('library_name: transformers', 'library_name: peft', 1) |
| elif kind == 'gguf': |
| content = content.replace('library_name: transformers', 'library_name: gguf', 1) |
| api.upload_file(repo_id=repo, repo_type=repo_type, path_or_fileobj=content.encode(), path_in_repo='README.md', commit_message='Document actual data, measurements, and limitations') |
| if kind == 'merged': |
| if video.exists(): |
| api.upload_file(repo_id=repo, path_or_fileobj=str(video), path_in_repo='VibeGame-launch-60s.mp4', |
| commit_message='Add measured 60-second model walkthrough') |
| api.upload_file(repo_id=repo, path_or_fileobj=sanitized_text(video_evidence.read_text()).encode(), |
| path_in_repo='VibeGame-launch-60s.demo.json', commit_message='Preserve video provenance') |
| for upload in uploads: |
| with tempfile.TemporaryDirectory(prefix='vibegame-public-evidence-') as temporary: |
| stage = Path(temporary) |
| stage_evaluation_upload(upload, stage) |
| api.upload_folder(repo_id=repo, repo_type='model', folder_path=stage, path_in_repo=upload['prefix'], |
| commit_message='Preserve separate experiment evidence and protocol receipts') |
| import subprocess |
| subprocess.run([sys.executable, str(ROOT / 'scripts/build_notebook.py')], check=True, cwd=ROOT) |
| with tempfile.TemporaryDirectory(prefix='vibegame-open-source-') as temporary: |
| stage = Path(temporary) |
| source_patterns = ['src/vibegame/*.py', 'scripts/*.py', 'tests/*.py', 'docs/*.md', 'requirements-colab.txt', 'pyproject.toml', 'LICENSE', 'NOTICE', 'README.md', 'notebooks/*.ipynb'] |
| for pattern in source_patterns: |
| for local in ROOT.glob(pattern): |
| if local.is_file(): |
| target = stage / local.relative_to(ROOT) |
| target.parent.mkdir(parents=True, exist_ok=True) |
| target.write_bytes(local.read_bytes()) |
| (stage / 'Modelfile').write_bytes((args.gguf / 'Modelfile').read_bytes()) |
| api.upload_folder(repo_id=repo, repo_type='model', folder_path=stage, path_in_repo='code', |
| commit_message='Release pipeline source, tests, notebook, and licenses') |
| if kind != 'dataset': |
| mapping = provider_mapping(api, repo) |
| report['provider_checks'][repo]['after'] = mapping |
| if mapping: |
| |
| |
| api.update_repo_settings(repo_id=repo, private=True) |
| raise RuntimeError(f'Provider mapping appeared for {repo}; release made private for investigation') |
| report.setdefault('uploaded', []).append(repo) |
| save_report() |
| |
| |
| api.promote() |
| report['status'] = 'artifacts_uploaded_account_reverification_required' |
| report['links'] = {kind: 'https://huggingface.co/' + ('datasets/' if kind == 'dataset' else '') + repo for kind, repo in repo_ids.items()} |
| |
| |
| try: |
| endpoints = api.list_inference_endpoints(namespace='*') |
| report['existing_inference_endpoints'] = [{'name': x.name, 'namespace': x.namespace, 'status': x.status} for x in endpoints] |
| except Exception: |
| report['existing_inference_endpoints'] = 'Not inspectable with this token; verify in account UI.' |
| save_report() |
| print(json.dumps({'status': report['status'], 'links': report['links'], 'provider_mappings': 'empty at checks', |
| 'account_settings': 'Reverify spend limit 0 and all providers disabled in account settings now.'}, indent=2)) |
| except Exception as exc: |
| report['status'] = 'failed' |
| |
| report['failure_type'] = type(exc).__name__ |
| save_report() |
| raise RuntimeError('Publication stopped. See the credential-free report and the local artifact checks; no token was persisted.') from None |
| finally: |
| token = None |
| api = None |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
