claude-turing 3.0.0 → 3.1.0

package/templates/scripts/sanity_checks.py

@@ -0,0 +1,503 @@
+#!/usr/bin/env python3
+"""Pre-training sanity checks for the autoresearch pipeline.
+
+Runs a battery of fast checks before committing to a full training run:
+initial loss validation, single-batch overfit, gradient flow, output
+validation, data pipeline check, and config consistency.
+
+Usage:
+    python scripts/sanity_checks.py
+    python scripts/sanity_checks.py --quick
+    python scripts/sanity_checks.py --verbose
+    python scripts/sanity_checks.py --json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+import numpy as np
+import yaml
+
+from scripts.turing_io import load_config
+
+DEFAULT_OVERFIT_STEPS = 50
+DEFAULT_OVERFIT_THRESHOLD = 0.1  # Loss should drop below this fraction of initial
+
+
+# --- Individual Checks ---
+
+
+def check_initial_loss(
+    initial_loss: float,
+    num_classes: int | None = None,
+    task_type: str = "classification",
+) -> dict:
+    """Check if initial loss matches theoretical expectation.
+
+    For classification with cross-entropy: expected = -log(1/num_classes).
+    """
+    check = {
+        "check": "initial_loss",
+        "severity": "high",
+        "initial_loss": round(initial_loss, 4),
+    }
+
+    if math.isnan(initial_loss) or math.isinf(initial_loss):
+        check["status"] = "fail"
+        check["reason"] = f"Initial loss is {initial_loss} — model is broken before training starts"
+        return check
+
+    if task_type == "classification" and num_classes and num_classes > 1:
+        expected = -math.log(1.0 / num_classes)
+        ratio = initial_loss / expected if expected > 0 else float("inf")
+        check["expected_loss"] = round(expected, 4)
+        check["ratio"] = round(ratio, 2)
+
+        if ratio > 3.0:
+            check["status"] = "fail"
+            check["reason"] = f"Initial loss {initial_loss:.4f} is {ratio:.1f}x expected ({expected:.4f}) — likely misconfigured loss function"
+        elif ratio > 2.0:
+            check["status"] = "warn"
+            check["reason"] = f"Initial loss {initial_loss:.4f} is {ratio:.1f}x expected ({expected:.4f}) — investigate"
+        else:
+            check["status"] = "pass"
+            check["reason"] = f"Initial loss {initial_loss:.4f} matches expected {expected:.4f} (ratio: {ratio:.2f})"
+    else:
+        # For regression, just check it's finite and reasonable
+        if initial_loss < 0:
+            check["status"] = "warn"
+            check["reason"] = f"Negative initial loss ({initial_loss:.4f}) — unusual for most loss functions"
+        else:
+            check["status"] = "pass"
+            check["reason"] = f"Initial loss {initial_loss:.4f} is finite and non-negative"
+
+    return check
+
+
+def check_single_batch_overfit(
+    loss_history: list[float],
+    threshold: float = DEFAULT_OVERFIT_THRESHOLD,
+) -> dict:
+    """Check if model can overfit a single batch.
+
+    If loss doesn't approach zero after N steps, something is broken.
+    """
+    check = {
+        "check": "single_batch_overfit",
+        "severity": "critical",
+    }
+
+    if not loss_history:
+        check["status"] = "skip"
+        check["reason"] = "No loss history provided"
+        return check
+
+    initial = loss_history[0]
+    final = loss_history[-1]
+    n_steps = len(loss_history)
+
+    if initial <= 0:
+        check["status"] = "warn"
+        check["reason"] = f"Initial loss is {initial:.4f} (non-positive), cannot assess overfit"
+        return check
+
+    reduction = 1 - (final / initial)
+
+    check["initial_loss"] = round(initial, 4)
+    check["final_loss"] = round(final, 4)
+    check["n_steps"] = n_steps
+    check["reduction"] = round(reduction, 4)
+
+    if any(math.isnan(l) for l in loss_history):
+        check["status"] = "fail"
+        check["reason"] = f"NaN in loss during overfit test — numerical instability"
+        return check
+
+    if reduction > 0.9:
+        check["status"] = "pass"
+        check["reason"] = f"Loss reduced by {reduction:.0%} in {n_steps} steps — model can memorize"
+    elif reduction > 0.5:
+        check["status"] = "warn"
+        check["reason"] = f"Loss reduced by only {reduction:.0%} — model is learning but slowly. Check learning rate."
+    else:
+        check["status"] = "fail"
+        check["reason"] = f"Loss stuck (reduced only {reduction:.0%} in {n_steps} steps) — model cannot memorize 1 batch. Check: architecture, learning rate, loss function"
+
+    return check
+
+
+def check_gradient_flow(
+    gradient_stats: list[dict],
+) -> dict:
+    """Check that gradients are non-zero and non-exploding for every parameter.
+
+    Args:
+        gradient_stats: List of {name, mean, max, min, std} per parameter group.
+    """
+    check = {
+        "check": "gradient_flow",
+        "severity": "high",
+    }
+
+    if not gradient_stats:
+        check["status"] = "skip"
+        check["reason"] = "No gradient statistics provided"
+        return check
+
+    dead_layers = []
+    exploding_layers = []
+    total = len(gradient_stats)
+
+    mean_grad = np.mean([abs(g.get("mean", 0)) for g in gradient_stats])
+
+    for g in gradient_stats:
+        name = g.get("name", "?")
+        grad_mean = abs(g.get("mean", 0))
+        grad_max = abs(g.get("max", 0))
+
+        if grad_mean == 0 and grad_max == 0:
+            dead_layers.append(name)
+        elif mean_grad > 0 and grad_max > 100 * mean_grad:
+            exploding_layers.append(name)
+
+    check["total_params"] = total
+    check["dead_layers"] = dead_layers
+    check["exploding_layers"] = exploding_layers
+
+    if dead_layers and exploding_layers:
+        check["status"] = "fail"
+        check["reason"] = f"{len(dead_layers)} dead layer(s) and {len(exploding_layers)} exploding layer(s)"
+    elif dead_layers:
+        check["status"] = "warn"
+        check["reason"] = f"{len(dead_layers)} dead layer(s) with zero gradients: {', '.join(dead_layers[:3])}"
+    elif exploding_layers:
+        check["status"] = "warn"
+        check["reason"] = f"{len(exploding_layers)} layer(s) with exploding gradients: {', '.join(exploding_layers[:3])}"
+    else:
+        check["status"] = "pass"
+        check["reason"] = f"All {total} parameter groups have non-zero, stable gradients"
+
+    return check
+
+
+def check_output_validation(
+    outputs: np.ndarray | list,
+    task_type: str = "classification",
+) -> dict:
+    """Check that model outputs are valid (non-NaN, non-constant, reasonable range)."""
+    check = {
+        "check": "output_validation",
+        "severity": "high",
+    }
+
+    arr = np.asarray(outputs, dtype=float)
+
+    if arr.size == 0:
+        check["status"] = "skip"
+        check["reason"] = "No outputs to validate"
+        return check
+
+    has_nan = bool(np.any(np.isnan(arr)))
+    has_inf = bool(np.any(np.isinf(arr)))
+    is_constant = bool(np.std(arr) == 0)
+    out_min = float(np.nanmin(arr))
+    out_max = float(np.nanmax(arr))
+
+    check["range"] = [round(out_min, 4), round(out_max, 4)]
+    check["has_nan"] = has_nan
+    check["has_inf"] = has_inf
+    check["is_constant"] = is_constant
+
+    issues = []
+    if has_nan:
+        issues.append("NaN values in outputs")
+    if has_inf:
+        issues.append("Inf values in outputs")
+    if is_constant:
+        issues.append("All outputs identical (constant predictions)")
+    if abs(out_max) > 100 and task_type == "classification":
+        issues.append(f"Extreme output range [{out_min:.1f}, {out_max:.1f}] — consider clamping")
+
+    if has_nan or has_inf:
+        check["status"] = "fail"
+        check["reason"] = "; ".join(issues)
+    elif is_constant:
+        check["status"] = "fail"
+        check["reason"] = "Constant predictions — model is not differentiating inputs"
+    elif issues:
+        check["status"] = "warn"
+        check["reason"] = "; ".join(issues)
+    else:
+        check["status"] = "pass"
+        check["reason"] = f"Outputs valid: range [{out_min:.4f}, {out_max:.4f}], no NaN/Inf"
+
+    return check
+
+
+def check_data_pipeline(
+    batch_shapes: dict | None = None,
+    has_nan: bool = False,
+    has_inf: bool = False,
+    loads_ok: bool = True,
+) -> dict:
+    """Check that the data pipeline produces valid batches."""
+    check = {
+        "check": "data_pipeline",
+        "severity": "critical",
+    }
+
+    if not loads_ok:
+        check["status"] = "fail"
+        check["reason"] = "Data pipeline failed to load first batch"
+        return check
+
+    issues = []
+    if has_nan:
+        issues.append("NaN values in input data")
+    if has_inf:
+        issues.append("Inf values in input data")
+
+    if batch_shapes:
+        check["shapes"] = batch_shapes
+
+    if issues:
+        check["status"] = "fail"
+        check["reason"] = "; ".join(issues)
+    elif batch_shapes:
+        shapes_str = ", ".join(f"{k}: {v}" for k, v in batch_shapes.items())
+        check["status"] = "pass"
+        check["reason"] = f"Batch loads, shapes correct ({shapes_str})"
+    else:
+        check["status"] = "pass"
+        check["reason"] = "Data pipeline functional"
+
+    return check
+
+
+def check_config_consistency(config: dict) -> dict:
+    """Check that config values are in reasonable ranges."""
+    check = {
+        "check": "config_consistency",
+        "severity": "medium",
+    }
+
+    issues = []
+    hyperparams = config.get("model", {}).get("hyperparams", {})
+
+    lr = hyperparams.get("learning_rate", hyperparams.get("lr"))
+    if lr is not None:
+        if lr > 1.0:
+            issues.append(f"Learning rate {lr} > 1.0 — unusually high")
+        elif lr < 1e-8:
+            issues.append(f"Learning rate {lr} < 1e-8 — effectively zero")
+
+    batch_size = hyperparams.get("batch_size")
+    if batch_size is not None:
+        if batch_size < 1:
+            issues.append(f"Batch size {batch_size} < 1 — invalid")
+        elif batch_size > 100000:
+            issues.append(f"Batch size {batch_size} > 100K — unusually large")
+
+    n_estimators = hyperparams.get("n_estimators")
+    if n_estimators is not None and n_estimators < 1:
+        issues.append(f"n_estimators {n_estimators} < 1 — invalid")
+
+    if issues:
+        check["status"] = "warn"
+        check["reason"] = "; ".join(issues)
+        check["issues"] = issues
+    else:
+        check["status"] = "pass"
+        check["reason"] = "Config values in reasonable ranges"
+
+    return check
+
+
+# --- Full Sanity Check ---
+
+
+def run_sanity_checks(
+    config_path: str = "config.yaml",
+    quick: bool = False,
+    initial_loss: float | None = None,
+    loss_history: list[float] | None = None,
+    gradient_stats: list[dict] | None = None,
+    outputs: list | None = None,
+    batch_shapes: dict | None = None,
+    data_has_nan: bool = False,
+    data_has_inf: bool = False,
+    data_loads_ok: bool = True,
+    num_classes: int | None = None,
+) -> dict:
+    """Run all sanity checks and produce a report.
+
+    In CLI mode, most inputs come from running a quick training probe.
+    In test mode, values are provided directly.
+    """
+    config = load_config(config_path)
+    task_type = config.get("task", {}).get("type", "classification")
+
+    checks = []
+
+    # Data pipeline
+    checks.append(check_data_pipeline(batch_shapes, data_has_nan, data_has_inf, data_loads_ok))
+
+    # Initial loss
+    if initial_loss is not None:
+        checks.append(check_initial_loss(initial_loss, num_classes, task_type))
+
+    # Gradient flow
+    if gradient_stats is not None:
+        checks.append(check_gradient_flow(gradient_stats))
+
+    # Single-batch overfit (skip in quick mode)
+    if not quick and loss_history is not None:
+        checks.append(check_single_batch_overfit(loss_history))
+
+    # Output validation
+    if outputs is not None:
+        checks.append(check_output_validation(outputs, task_type))
+
+    # Config consistency
+    checks.append(check_config_consistency(config))
+
+    # Compute verdict
+    n_pass = sum(1 for c in checks if c["status"] == "pass")
+    n_fail = sum(1 for c in checks if c["status"] == "fail")
+    n_warn = sum(1 for c in checks if c["status"] == "warn")
+    n_skip = sum(1 for c in checks if c["status"] == "skip")
+
+    if n_fail > 0:
+        verdict = "fail"
+    elif n_warn > 2:
+        verdict = "warn"
+    elif n_warn > 0:
+        verdict = "pass_with_warnings"
+    else:
+        verdict = "pass"
+
+    return {
+        "checked_at": datetime.now(timezone.utc).isoformat(),
+        "quick_mode": quick,
+        "task_type": task_type,
+        "checks": checks,
+        "score": {
+            "pass": n_pass,
+            "fail": n_fail,
+            "warn": n_warn,
+            "skip": n_skip,
+            "total": len(checks),
+        },
+        "verdict": verdict,
+    }
+
+
+# --- Report Formatting ---
+
+
+def save_sanity_report(report: dict, output_dir: str = "experiments/sanity") -> Path:
+    """Save sanity report to YAML."""
+    out_path = Path(output_dir)
+    out_path.mkdir(parents=True, exist_ok=True)
+
+    timestamp = datetime.now(timezone.utc).strftime("%Y%m%d-%H%M%S")
+    filepath = out_path / f"sanity-{timestamp}.yaml"
+
+    with open(filepath, "w") as f:
+        yaml.dump(report, f, default_flow_style=False, sort_keys=False)
+
+    return filepath
+
+
+def format_sanity_report(report: dict) -> str:
+    """Format sanity report as markdown."""
+    if "error" in report:
+        return f"ERROR: {report['error']}"
+
+    verdict = report.get("verdict", "?")
+    score = report.get("score", {})
+    quick = report.get("quick_mode", False)
+
+    verdict_labels = {
+        "pass": "PASS — Safe to proceed with training",
+        "pass_with_warnings": "PASS (with warnings) — Review before training",
+        "warn": "WARNINGS — Multiple issues detected",
+        "fail": "FAIL — Do not proceed to full training",
+    }
+
+    lines = [
+        "# Sanity Check Report",
+        "",
+        f"*Checked {report.get('checked_at', 'N/A')[:19]}*",
+        f"*Mode: {'quick' if quick else 'full'}*",
+        "",
+        f"**{verdict_labels.get(verdict, verdict.upper())}**",
+        "",
+    ]
+
+    status_markers = {"pass": "PASS", "fail": "FAIL", "warn": "WARN", "skip": "SKIP"}
+
+    for c in report.get("checks", []):
+        status = c.get("status", "?")
+        marker = status_markers.get(status, status.upper())
+        lines.append(f"- **[{marker}]** {c.get('check', '?')}: {c.get('reason', 'N/A')}")
+
+    lines.extend([
+        "",
+        f"**Score:** {score.get('pass', 0)}/{score.get('total', 0)} pass, "
+        f"{score.get('fail', 0)} fail, {score.get('warn', 0)} warn",
+    ])
+
+    return "\n".join(lines)
+
+
+def main() -> None:
+    """CLI entry point."""
+    parser = argparse.ArgumentParser(
+        description="Pre-training sanity checks",
+    )
+    parser.add_argument(
+        "--config", default="config.yaml",
+        help="Path to config.yaml",
+    )
+    parser.add_argument(
+        "--quick", action="store_true",
+        help="Quick mode: skip single-batch overfit test",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="Show detailed check output",
+    )
+    parser.add_argument(
+        "--json", action="store_true",
+        help="Output raw JSON instead of formatted report",
+    )
+    args = parser.parse_args()
+
+    # In CLI mode, we'd run actual probes. For now, report with config check only.
+    report = run_sanity_checks(
+        config_path=args.config,
+        quick=args.quick,
+    )
+
+    if "error" not in report:
+        filepath = save_sanity_report(report)
+        print(f"Saved to {filepath}", file=sys.stderr)
+
+    if args.json:
+        print(json.dumps(report, indent=2, default=str))
+    else:
+        print(format_sanity_report(report))
+
+    if report.get("verdict") == "fail":
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

package/templates/scripts/scaffold.py

@@ -118,6 +118,9 @@ TEMPLATE_DIRS = {
         "model_distiller.py",
         "knowledge_transfer.py",
         "methodology_audit.py",
+        "sanity_checks.py",
+        "generate_baselines.py",
+        "leakage_detector.py",
     ],
     "tests": ["__init__.py", "conftest.py"],
 }
@@ -148,6 +151,9 @@ DIRECTORIES_TO_CREATE = [
     "experiments/distillations",
     "experiments/transfers",
     "experiments/audits",
+    "experiments/sanity",
+    "experiments/baselines",
+    "experiments/leakage",
     "experiments/logs",
     "models/best",
     "models/archive",