WO56: attic the superseded framework stack + card frontmatter

.github/workflows/release.yml

@@ -60,9 +60,9 @@ jobs:
         run: |
           mkdir -p models/$TASK
           if [ "$VARIANT" = "fp32" ]; then
-            python -m src.cli export --task $TASK --data-root data --out-root models/$TASK
+            PYTHONPATH=attic python -m cli export --task $TASK --data-root data --out-root models/$TASK
           else
-            python -m src.cli export --task $TASK --variant $VARIANT --data-root data --out-root models/$TASK
+            PYTHONPATH=attic python -m cli export --task $TASK --variant $VARIANT --data-root data --out-root models/$TASK
           fi
       - name: Compute SHA256
         env:
@@ -94,7 +94,7 @@ jobs:
         env:
           TASK: ${{ needs.parse-tag.outputs.task }}
         run: |
-          python -m src.cli evaluate --task $TASK --data-root data --out-root models \
+          PYTHONPATH=attic python -m cli evaluate --task $TASK --data-root data --out-root models \
             > $TASK-benchmarks.json
       - uses: actions/upload-artifact@v4
         with:

.github/workflows/test.yml

@@ -39,10 +39,10 @@ jobs:
         with:
           python-version: "3.11"
       - run: pip install -e ".[dev,export]"
-      - name: CLI smoke test
+      - name: CLI smoke test (isxml)
         run: |
-          PYTHONPATH=src python -m src.cli list
-          test "$(PYTHONPATH=src python -m src.cli list | jq 'length')" -ge 3
+          PYTHONPATH=src isxml --help
+          PYTHONPATH=src isxml validate --help
 
   export-fixture:
     runs-on: ubuntu-latest

attic/gpu/probe_heb_parity.py

@@ -0,0 +1,79 @@
+"""Probe: why did heb-diac-1.1-fp32 fail the parity gate?
+
+Runs the exact parity harness (torch reference vs ONNX kv decode) but
+PRINTS the report numbers and sample outputs instead of raising.
+
+    modal run --detach src/gpu/probe_heb_parity.py
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import modal
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+
+IMAGE = (
+    modal.Image.debian_slim(python_version="3.11")
+    .pip_install("torch==2.12.1", "transformers==5.14.1", "onnx==1.22.0",
+                 "onnxruntime==1.23.2", "pyyaml>=6.0")
+    .add_local_dir(str(REPO_ROOT), "/root/interscript-ml", copy=True)
+    .workdir("/root/interscript-ml")
+)
+
+CHECKPOINT_VOLUMES = {"/volumes/rababa-checkpoints": modal.Volume.from_name("rababa-checkpoints")}
+DATASET_VOLUMES = {"/datasets/rababa": modal.Volume.from_name("rababa-datasets")}
+MODELS_VOLUME = modal.Volume.from_name("secryst-models")
+
+app = modal.App("probe-heb-parity", image=IMAGE)
+
+
+@app.function(
+    cpu=8,
+    memory=32 * 1024,
+    timeout=2 * 3600,
+    volumes={**CHECKPOINT_VOLUMES, **DATASET_VOLUMES, "/outputs": MODELS_VOLUME},
+)
+def probe() -> dict:
+    import json
+    import sys
+
+    sys.path.insert(0, "/root/interscript-ml/src")
+
+    for v in list(CHECKPOINT_VOLUMES.values()) + list(DATASET_VOLUMES.values()) + [MODELS_VOLUME]:
+        v.reload()
+
+    from imf.export import load_byte_seq2seq
+    from imf.parity import reference_decode, run_parity
+
+    ckpt = Path("/volumes/rababa-checkpoints/rababa_hebrew/run-s46-phonikud-plus/run-002-gold-ft/best")
+    zip_path = Path("/outputs/imf/heb-diac/heb-diac-1.1-fp32.zip")
+
+    model = load_byte_seq2seq(ckpt)
+
+    pairs = []
+    for line in Path("/datasets/rababa/nakdimon/test-imf.jsonl").read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        row = json.loads(line)
+        pairs.append((row.get("input", row.get("src", "")), row.get("target", row.get("tgt", ""))))
+    pairs = [(s, t) for s, t in pairs if s][:40]
+    print(f"[probe] {len(pairs)} pairs", flush=True)
+
+    reference = reference_decode(model, [s for s, _ in pairs], max_len=128)
+    for i in [0, 1, 2]:
+        print(f"[ref {i}] {reference[i][:90]!r}", flush=True)
+
+    report = run_parity(model, zip_path, pairs, max_len=128, reference=reference)
+    print(f"REPORT: samples={report.samples} cer_ref={report.cer_reference}pp "
+          f"cer_onnx={report.cer_onnx}pp delta={report.cer_delta}pp "
+          f"mismatches={report.token_mismatches} passed={report.passed}", flush=True)
+    return {"samples": report.samples, "cer_ref": report.cer_reference,
+            "cer_onnx": report.cer_onnx, "delta": report.cer_delta,
+            "mismatches": report.token_mismatches}
+
+
+@app.local_entrypoint()
+def main():
+    print(probe.remote())

cards/fas-diac-base-1.0.md

@@ -1,3 +1,15 @@
+---
+license: bsd-3-clause
+language:
+  - fa
+tags:
+  - diacritization
+  - byt5
+  - imf
+  - persian
+metrics:
+  - cer
+---
 # fas-diac-base-1.0
 
 Persian diacritization, base tier (run-077): byt5-small seq2seq on

cards/urd-diac-base-1.0.md

@@ -1,3 +1,17 @@
+---
+license: other
+license_name: bsd-3-clause-with-data-dependency
+license_link: https://github.com/interscript/interscript-models/blob/main/cards/urd-diac-base-1.0.md
+language:
+  - ur
+tags:
+  - diacritization
+  - byt5
+  - imf
+  - urdu
+metrics:
+  - cer
+---
 # urd-diac-base-1.0
 
 Urdu diacritization, base tier (run-066): byt5-base seq2seq on the

cards/urd-g2p-base-1.0.md

@@ -1,3 +1,18 @@
+---
+license: other
+license_name: research-only-non-commercial-data
+license_link: https://github.com/interscript/interscript-models/blob/main/cards/urd-g2p-base-1.0.md
+language:
+  - ur
+tags:
+  - g2p
+  - byt5
+  - imf
+  - urdu
+  - research-only
+metrics:
+  - cer
+---
 # urd-g2p-base-1.0
 
 Urdu grapheme-to-phoneme, base tier (run-059): byt5-base seq2seq on

pyproject.toml

@@ -58,7 +58,7 @@ isxml = "pipeline.cli:main"
 
 [tool.setuptools.packages.find]
 where = ["src"]
-include = ["framework*", "tasks*", "imf*", "sadeedbench*", "pipeline*"]
+include = ["imf*", "sadeedbench*", "pipeline*"]
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

src/imf/parity.py

@@ -19,7 +19,6 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from framework.evaluator import char_error_rate
 from imf.export import (
     BYTE_OFFSET,
     EOS_ID,
@@ -29,6 +28,7 @@
     onnx_greedy_kv,
 )
 from imf.schema import ModelMetadata, Parity
+from imf.text_metrics import char_error_rate
 
 
 @dataclass(frozen=True)

src/imf/text_metrics.py

@@ -0,0 +1,34 @@
+"""Character-level edit metrics for the IMF parity gate.
+
+Inlined from the retired framework module (WO56 attic move): the gate
+needs CER and nothing else. Same definitions, one dependency-free
+home.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+
+def edit_distance(a: Sequence[object], b: Sequence[object]) -> int:
+    """Levenshtein distance."""
+    if a == b:
+        return 0
+    if not a:
+        return len(b)
+    if not b:
+        return len(a)
+    prev = list(range(len(b) + 1))
+    for i, ca in enumerate(a, 1):
+        cur = [i]
+        for j, cb in enumerate(b, 1):
+            cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb)))
+        prev = cur
+    return prev[-1]
+
+
+def char_error_rate(pred: str, gold: str) -> float:
+    """CER = edit_distance / len(gold)."""
+    if not gold:
+        return 0.0 if not pred else 1.0
+    return edit_distance(list(pred), list(gold)) / len(gold)

src/pipeline/cli.py

@@ -100,6 +100,36 @@ def cmd_conformance(args: argparse.Namespace) -> int:
     return 0 if res.passed else 1
 
 
+def cmd_housekeeper(args: argparse.Namespace) -> int:
+    """Live bucket inventory + prune plan. Classify-only: this verb
+    cannot delete anything (WO56 law) — the plan it prints is the
+    input to a separate, owner-authorized execution step."""
+    import os
+
+    from huggingface_hub import HfFileSystem
+
+    from pipeline.housekeeper import inventory, prune_plan
+
+    fs = HfFileSystem(token=os.environ.get("HF_TOKEN"))
+    runs = {}
+    for f in fs.ls(args.bucket):
+        name = f["name"].rsplit("/", 1)[-1]
+        if name.startswith(args.prefix):
+            runs[name] = fs.ls(f["name"])
+    inv = inventory(runs)
+    for r in inv:
+        status = "complete" if r.complete else "IN-FLIGHT"
+        notes = ("; ".join(r.notes)) if r.notes else ""
+        print(f"{r.name:40s} {status:10s} {r.files:4d} files "
+              f"{r.bytes / 1e9:6.2f} GB  {notes}")
+    plan = prune_plan(inv)
+    print(f"\nprune plan: {len(plan)} candidate(s) "
+          "(DRY-RUN — execution is owner-authorized, this verb cannot delete):")
+    for p in plan:
+        print(" ", p)
+    return 0
+
+
 def main(argv: list[str] | None = None) -> int:
     ap = argparse.ArgumentParser(prog="isxml")
     sub = ap.add_subparsers(dest="cmd", required=True)
@@ -130,6 +160,12 @@ def main(argv: list[str] | None = None) -> int:
     c.add_argument("--zip", required=True, type=Path)
     c.set_defaults(fn=cmd_conformance)
 
+    h = sub.add_parser("housekeeper",
+                       help="bucket inventory + prune plan (dry-run only)")
+    h.add_argument("--bucket", default="buckets/Interscript/isx-training")
+    h.add_argument("--prefix", default="run-")
+    h.set_defaults(fn=cmd_housekeeper)
+
     args = ap.parse_args(argv)
     return args.fn(args)