from dataclasses import dataclass import fcntl import json from pathlib import Path import shutil import time from huggingface_hub import CommitOperationAdd, HfApi from omegaconf import OmegaConf from rich import print from refined_v4 import publish as shared from .data import ROOT, TASKS REPO = 'davidheineman/tiny-rl-refined-math-v1' STAGES = ( ('full', '', 16777216, '', 'sft', 'rl_envs'), ('pilot-2m', 'pilot-2m', 2097152, 'pilot-2m/', 'sft_2m', 'rl_envs_2m'), ('aligned-2m', 'iterations/aligned-v1', 2097152, 'aligned-2m/', 'sft_aligned_2m', 'rl_envs_aligned_2m'), ('canonical-2m', 'iterations/canonical-v1', 2097152, 'canonical-2m/', 'sft_canonical_2m', 'rl_envs_canonical_2m'), ('bridge-2m', 'iterations/bridge-v1', 2097152, 'bridge-2m/', 'sft_bridge_2m', 'rl_envs_bridge_2m'), ('smallcalc-2m', 'iterations/smallcalc-v1', 2097152, 'smallcalc-2m/', 'sft_smallcalc_2m', 'rl_envs_smallcalc_2m'), ('auxiliary-2m', 'iterations/aux-smallmath-v1', 2097152, 'auxiliary-2m/', 'sft_auxiliary_2m', 'rl_envs_auxiliary_2m'), ('visible-2m', 'iterations/gsm-visible-v2', 2097152, 'visible-2m/', 'sft_visible_2m', 'rl_envs_visible_2m'), ('final-step-2m', 'iterations/gsm-final-step-v3', 2097152, 'final-step-2m/', 'sft_final_step_2m', 'rl_envs_final_step_2m'), ('auxiliary-full', 'curricula/auxiliary-full-v1', 16777216, 'auxiliary-full/', 'sft_auxiliary', 'rl_envs_auxiliary'), ('bridge-full', 'curricula/bridge-full-v1', 16777216, 'bridge-full/', 'sft_bridge', 'rl_envs_bridge'), ('visible-full', 'curricula/gsm-visible-full-v2', 16777216, 'visible-full/', 'sft_visible', 'rl_envs_visible'), ('final-step-full', 'curricula/gsm-final-step-full-v3', 16777216, 'final-step-full/', 'sft_final_step', 'rl_envs_final_step')) @dataclass class Config: root: str = str(ROOT) repo: str = REPO action: str = 'upload' interval: int = 60 def card(states: dict, repo: str, verification: dict) -> str: lines = ['---', 'configs:'] for stage, source, budget, prefix, config, environment_config in STAGES: if any(state.get('sft_fingerprint') for state in states.get(stage, {}).values()): lines += [f'- config_name: {config}', ' data_files:', ' - split: train', f' path: {prefix}data/sft/*/*.parquet'] if states.get(stage): lines += [f'- config_name: {environment_config}', ' data_files:'] for split in ('train', 'validation', 'test'): lines += [f' - split: {split}', f' path: {prefix}data/rl_envs/{split}/*/*.parquet'] lines += ['---', '', '# Tiny RL math-v1', '', 'Repaired GSM8k, Partial GSM8k, Arithmetic and Arithmetic Simple datasets. ' 'The full SFT budget is 16,777,216 predictions per task; pilots contain 2,097,152. ' 'SFT documents contain 2–4 native examples, without system prompts or chat control tokens.', '', 'GSM uses original questions and calculator annotations from openai/gsm8k at pinned revision ' '740312add88f781978c0658806c59bc2815b9866. Each equation is checked exactly. ' 'The API supplies only short generic phrases from a closed vocabulary; it cannot change facts, equations, or answers. ' 'Original GSM test questions stay held out. Source exclusions and duplicate handling are recorded in source-audit.json.', '', 'Arithmetic has one or two operations. Its SFT mixture is 40% tiny addition/subtraction with 0–9 inputs, ' '30% addition/subtraction with 0–20 inputs, 15% multiplication with 0–9 inputs, and 15% two-step small expressions. ' 'Answers are bounded between -18 and 81. Repeated small problems are intentional; diversity is not obtained by increasing problem size.', '', 'Choose the missing-token count uniformly from 1 through 16, then choose a complexity band and an eligible source. ' 'Renormalize mixture weights over the bands available at that length. ' 'Short references support low hints and are retained. The environment configuration exports only eligible source/hint pairs; ' 'not every source supports all 16 cuts. `provenance` records difficulty, operation count, answer width, and immutable equations.', '', 'All exported eligible gold suffixes pass the numeric oracle. Source groups are disjoint across ' 'SFT/RL/calibration/development/test. Demonstrations use SFT groups. Native templates and source records are included ' 'under raw/, provenance/, styles.json, and code/. Reconstruct a row from its source identity and style_id; ' 'style_id=-1 selects DEFAULT_STYLE in code/refined_math/data.py. ' 'the SFT text column is the actual training document.', '', 'Use sft / rl_envs for full-budget data and sft_2m / rl_envs_2m for pilot data. ' 'RL train/validation/test correspond to RL/dev/test sources. JSON-valued columns are serialized JSON strings.', '', '`sft_aligned_2m` is a later controlled revision: about half its examples use the exact frozen ' 'environment wording, and half use the original API wording. It retains the same source folds and exact 2M budget. ' 'Its matching environment config is rl_envs_aligned_2m; the environment assets are unchanged.', '', '`sft_canonical_2m` uses only the frozen environment wording. `sft_bridge_2m` uses ' 'approximately 75% computation subtasks and 25% original word problems for GSM. ' 'A subtask asks for the final verified equation from an SFT-only source; it retains ' 'the original source group, answer and question provenance. Arithmetic remains canonical. ' 'No held-out equations are used to produce SFT. Matching `rl_envs_canonical_2m` and ' '`rl_envs_bridge_2m` retain the original word-problem environments byte-for-byte. ' 'The styles and alignment audits record the exact mixtures. See code/refined_math/aligned.py.', '', '`sft_auxiliary_2m` uses about 75% bounded arithmetic examples in each GSM variant’s ' 'native layout and 25% original GSM word problems. Auxiliary arithmetic comes only ' 'from the independent arithmetic SFT fold, disjoint from every held-out source fold. ' 'Auxiliary identities and source domains remain explicit in provenance. Arithmetic tasks ' 'remain canonical. `rl_envs_auxiliary_2m` retains the original GSM word problems and all ' 'source/hint cuts; this curriculum changes SFT only.', '', '`sft_visible_2m` / `rl_envs_visible_2m` define the separate visible-calculation GSM revision. ' 'Original GSM questions retain the verified final calculation in the question context at every ' '1–16-token cut. Easy is addition/subtraction of integers in 0–20; standard has up to two ' 'operations on integer literals of magnitude at most 100 with integer result at most 1,000; ' 'hard retains the other original calculations. RL samples difficulty bands 70/20/10 after ' 'choosing the hint uniformly; evaluation uses equal bands. Literal final expressions are ' 'excluded because supplying them would directly give the answer. Source exclusions, exact ' 'budgets and protected-context checks are recorded in context-audit.json and source-audit.json. ' 'These scores describe a different hinted input from the original full-GSM challenge. ' 'Native layouts and multiple examples per document are retained. The 16M counterparts ' 'are sft_visible / rl_envs_visible; sft_auxiliary and sft_bridge provide the full earlier curricula.', '', '`sft_final_step_2m` / `rl_envs_final_step_2m` retain the visible calculation and the same ' 'GSM question folds and difficulty range, with a separate reference revision. References ' 'explain the supplied final operation in one sentence followed by the equation and answer, ' 'rather than a variable chain of intermediate steps. Operators use consistent spacing. ' 'This makes the stopping point explicit at longer cuts. This is an experimental task ' 'revision; matched learning results are listed below. Its full-budget counterparts ' 'are sft_final_step / rl_envs_final_step when ready. Reconstruct visible/final-step rows ' 'with code/refined_math/visible.py and the exported styles.json; rebuilding final-step ' 'data requires final_step_reference=true. See its context-audit.json.', '', '`sft_smallcalc_2m` is an isolated curriculum control using the original word problems ' 'and canonical wording. Its GSM source sampler draws 80% of examples from training sources ' 'whose final equation is a single addition/subtraction with integer operands 0..20, ' 'and 20% from the original mixture. The realized small-calculation fraction is slightly ' 'higher because the broad mixture also contains these problems. No questions or equations ' 'are rewritten. Arithmetic uses the ordinary canonical corpus. ' '`rl_envs_smallcalc_2m` retains the same complete word-problem pools; there is no ' 'easier evaluation subset. The training results are still being investigated.', '', 'The verified 32M Arithmetic Simple run uses the original sft_2m data, 512-token SFT batches, ' 'Q/K RMS 0.1, undamped Q/K updates, one-third RL rates, KL 0.1, and 65,536 RL samples. ' 'On 256 held-out source groups it scores 60.1% after SFT and 62.7% after RL. ' 'The paired 95% interval for the gain is +0.5 to +4.8 percentage points. ' 'Verification files are under verification/. GSM accuracy remains low and is still under investigation.', '', '| Held-out comparison | SFT pass@1 | RL pass@1 | Gain (pp) | Paired 95% interval (pp) | Questions |', '|---|---:|---:|---:|---:|---:|'] for name, result in sorted(verification.items()): low, high = result['uplift_ci95'] lines.append(f'| {name} | {result["sft_pass1"]:.2%} | {result["rl_pass1"]:.2%} | ' f'{100 * result["uplift"]:+.2f} | [{100 * low:+.2f}, {100 * high:+.2f}] | {result["queries"]} |') lines += ['', '| Stage | Task | RL sources | Dev sources | Test sources | SFT tokens |', '|---|---|---:|---:|---:|---:|'] for stage, tasks in states.items(): for task, state in sorted(tasks.items()): counts = state['environment_counts'] lines.append(f'| {stage} | {task} | {counts["train"]["source_rows"]} | ' f'{counts["validation"]["source_rows"]} | {counts["test"]["source_rows"]} | ' f'{state.get("sft_budget_tokens", "pending")} |') return '\n'.join(lines) + '\n' def materialize_ready_data(settings, task: str): original = shared.source_state def source_state(folder: Path, selected_task: str): environments, supervised = original(folder, selected_task) hold_path = folder / 'holds.json' hold = shared.read_json(hold_path).get(selected_task, {}) if hold_path.exists() else {} manifest_path = folder / 'sft' / selected_task / 'manifest.json' training_hold = str(hold.get('reason', '')).startswith('2M iterative debugging:') if supervised is None and training_hold and not hold.get('pending_revision') and manifest_path.exists(): candidate = shared.read_json(manifest_path) if (candidate.get('status') == 'sft_ready' and candidate.get('fresh_sft') is True and candidate.get('generation_scope') == 'verified_math'): supervised = candidate return environments, supervised shared.source_state = source_state try: return shared.materialize_task(settings, task) finally: shared.source_state = original def upload(config: Config): root = Path(config.root) stage = root / 'publication' stage.mkdir(parents=True, exist_ok=True) receipt = stage / 'published.json' previous = shared.read_json(receipt) if receipt.exists() else dict(files={}, states={}) states, files = {}, {} for label, relative, budget, prefix, sft_config, environment_config in STAGES: source = root / relative states[label] = {} if relative.startswith('curricula/') and not any((source / name).exists() for name in ('alignment-audit.json', 'context-audit.json')): continue settings = shared.Config(root=str(source), tasks=list(TASKS), tokens_per_task=budget) for task in TASKS: manifest = source / 'sft' / task / 'manifest.json' if not manifest.exists(): continue state = materialize_ready_data(settings, task) if state is None: continue states[label][task] = state for metadata in state['env_files'] + state.get('sft_files', []) + state.get('sft_provenance_files', []): path = source / 'publish' / metadata['path'] files[prefix + metadata['path']] = path files[prefix + 'manifests/' + task + '.json'] = source / 'publish/manifests' / (task + '.json') files[prefix + 'raw_sft/' + task + '.jsonl.gz'] = source / 'sft' / task / 'train.jsonl.gz' for name in ('styles.json', 'alignment-audit.json', 'context-audit.json', 'source-audit.json', 'pilot-prefix-audit.json'): if (source / name).exists(): files[prefix + name] = source / name for name in ('styles.json', 'source-audit.json'): if (root / name).exists(): files[name] = root / name failure_analysis = root / 'report-cache/failure-analysis.json' if failure_analysis.exists(): files['audits/failure-analysis.json'] = failure_analysis corpus_audit = root / 'report-cache/data-audit.json' if corpus_audit.exists(): files['data-audit.json'] = corpus_audit hint_probe = root / 'report-cache/final-step-hint1-easy-comparison.json' if hint_probe.exists(): files['audits/final-step-hint1-easy-comparison.json'] = hint_probe first_probe = root / 'iterations/gsm-final-step-v3/runs/single-final-calculation/8m/gsm8k/diagnostics' for name in ('hint1-easy-sft-cpu.json', 'hint1-easy-sft-cpu-protocol.json', 'hint1-easy-rl-cpu.json', 'hint1-easy-rl-cpu-protocol.json'): if (first_probe / name).exists(): files['audits/final-step-8m-' + name] = first_probe / name verification = {} for folder in sorted((root / 'verification').iterdir()): if not (folder / 'result.json').exists(): continue verification[folder.name] = shared.read_json(folder / 'result.json') for name in ('protocol.json', 'result.json', 'sampler-comparison.json', 'comparison-reused-eval.json'): if (folder / name).exists(): files['verification/' + folder.name + '/' + name] = folder / name for suite in ('aux-smallmath-v1', 'rl-smallcalc-v1', 'full-arithmetic-v1', 'gsm-visible-v2', 'gsm-final-step-v3', 'gsm-final-step-full-v3'): plan = root / 'iterations' / suite / 'plan.json' if plan.exists(): files['protocols/' + suite + '/plan.json'] = plan for case in shared.read_json(plan)['cases']: for name in ('rl-curriculum.json', 'reused-sft.json'): path = Path(case['output']) / name if path.exists(): files['protocols/' + suite + '/' + case['task'] + '-' + name] = path audit = root / 'report-cache/data-audit.json' if audit.exists(): files['audits/data-audit.json'] = audit if not any(states.values()): return previous for path in Path(__file__).parent.glob('*.py'): files['code/refined_math/' + path.name] = path readme = stage / 'README.md' readme.write_text(card(states, config.repo, verification)) files['README.md'] = readme changed = {name: path for name, path in files.items() if previous['files'].get(name) != shared.digest(path)} if not changed: return previous api = HfApi() api.create_repo(config.repo, repo_type='dataset', private=False, exist_ok=True) repository = api.repo_info(config.repo, repo_type='dataset') operations = [CommitOperationAdd(path_in_repo=name, path_or_fileobj=str(path)) for name, path in changed.items()] commit = api.create_commit(config.repo, repo_type='dataset', operations=operations, commit_message='Publish coherent math-v1 datasets', parent_commit=repository.sha, num_threads=4) remote = api.get_paths_info(config.repo, list(changed), repo_type='dataset', revision=commit.oid) by_path = {info.path: info for info in remote} for name, path in changed.items(): if name not in by_path or not shared.remote_matches(by_path[name], path): raise RuntimeError('Remote hash/size verification failed: ' + name) result = dict(repo=config.repo, commit=commit.oid, states=states, updated=time.time(), files={name: shared.digest(path) for name, path in files.items()}) shared.atomic_json(receipt, result) print(f'[green]Public data uploaded and hash-verified: https://huggingface.co/datasets/{config.repo}[/green]') return result def run(config: Config): root = Path(config.root) (root / 'publication').mkdir(parents=True, exist_ok=True) with (root / 'publication/publisher.lock').open('a') as lock: fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB) while True: result = upload(config) complete = all(len(result.get('states', {}).get(stage, {})) == 4 for stage in ('full', 'pilot-2m')) if config.action != 'watch' or complete: return time.sleep(config.interval) if __name__ == '__main__': run(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config), OmegaConf.from_cli())))