⭐ 1. Do all backends have a tagset?¶

Every backend will have a tagset, but not in the same way and not for the same purpose.

✔ CRF backend¶

Yes — CRF must have a tagset because it predicts BIO tags:

["B", "I"]

This is part of the model definition.

✔ Rule backend¶

No — rule backend does not use tags.
But it does have a “feature set” (onsets, nuclei, codas, rules).

✔ FST backend¶

No tagset — but it has states, transitions, etc.

✔ Transformer backend¶

No tagset — but it has layers, heads, hidden size.

✔ BiLSTM backend¶

If it predicts BIO tags, then yes — it will have a tagset.

⭐ 2. So how do we unify metadata across all backends?¶

We do NOT force every backend to have a tagset.
Instead:

• CRF backend → features.tagset
• Rule backend → features.ruleset_summary
• FST backend → features.fst_states
• Transformer backend → features.num_layers, etc.

This is exactly why UMS has:

"features": { ... backend-specific ... }

Each backend fills in only what applies.

⭐ 3. CRF backend implementation (UMS‑compliant)¶

Here is the correct UMS implementation for CRF backend, using existing interface.

✔ Add backend identity¶

✔ Add feature metadata¶

✔ Add UMS wrapper¶

✔ Keep existing capabilities block¶

class CRFBackend(BaseSyllabifier):
    backend_name = "crf"
    backend_type = "statistical"
    backend_version = "1.0.0"

    def _feature_metadata(self):
        return {
            "tagset": ["B", "I"],
            "num_features": len(self.feature_templates),
            "feature_templates": self.feature_templates,
            "model_path": str(self.model_path),
        }

    def get_metadata(self) -> Dict[str, Any]:
        return {
            "backend_type": "crf",
            "version": self.backend_version,
            "capabilities": {
                "supports_confidence": True,
                "supports_batch": True,
                "supports_gpu": False,
                "supports_cpu": True,
            },
            "ums": self._base_ums(),
        }

This is fully correct and matches backend interface.

⭐ 4. Why this is the correct design¶

✔ CRF backend has a tagset¶

Because it predicts BIO tags.

✔ Rule backend does NOT have a tagset¶

Because it does not predict tags.

✔ UMS allows backend‑specific features¶

Without forcing irrelevant fields.

✔ include_metadata=True now matters¶

Because metadata is only attached when requested.

⭐ 5. What the CRF metadata will look like¶

When a user calls:

zs.syllabify("itna", model="crf", return_metadata=True)

They will see:

raw["metadata"] = {
    "backend_type": "crf",
    "version": "1.0.0",
    "capabilities": {...},
    "ums": {
        "backend": {...},
        "model": {...},
        "features": {
            "tagset": ["B", "I"],
            "num_features": 128,
            "feature_templates": [...],
            "model_path": ".../crf_syllabifier.joblib"
        },
        "runtime": {...}
    }
}