Indic NER checkpoint compatibility¶

OpenMed's optional Indic NER adapter accepts only a user-selected Hugging Face token-classification checkpoint. It does not bundle weights, select a public repository, or contact a model host when OPENMED_INDIC_NER_MODEL is unset.

Install the optional runtime before loading a checkpoint:

pip install "openmed[hf]" torch

Then use an existing local directory or an explicitly approved repository:

from openmed.ner.families import load_indic_ner_adapter

local_adapter = load_indic_ner_adapter(
    "/srv/models/indic-ner",
    local_files_only=True,
)

remote_adapter = load_indic_ner_adapter(
    "your-org/approved-indic-ner",
    revision="reviewed-revision",
)

Existing paths are forced into local-only loading even if local_files_only=False. Repository resolution is possible only when a caller passes a repository identifier directly or sets OPENMED_INDIC_NER_MODEL. Remote code remains disabled in both layouts.

Compatibility matrix¶

A compatible map must contain an outside O label and semantic labels for all three CoNLL entity classes: PER, LOC, and ORG. The map must match config.num_labels when that value is present, and each model logit row must have the same width.

Unsupported contracts¶

The adapter fails closed for:

• slow tokenizers or tokenizers without exact character offsets;
• negative, reversed, out-of-bounds, overlapping, or backward offsets;
• generic LABEL_0-style maps without entity semantics;
• missing PER, LOC, or ORG labels, duplicate/non-contiguous indices, and conflicting label maps;
• CRF or custom heads that do not return token-aligned logits;
• checkpoints that require trust_remote_code=True;
• custom preprocessing whose offsets do not refer to the original Python string.

Compatibility exceptions contain only stable contract descriptions. They do not include the input string, checkpoint error detail, token surfaces, or credentials. Predictions serialize only offsets, canonical labels, and confidence values.

Synthetic robustness release gate¶

The committed Naamapadam-style evaluation uses synthetic records only. For each of the 11 supported Indic languages, it retains one acceptance baseline and adds deterministic cases for combining marks, punctuation-adjacent spans, repeated entity surfaces, code mixing, and Latin/native script boundaries. No Naamapadam corpus row or model weight is bundled.

A checkpoint is release-eligible only when all of these thresholds pass:

In the aggregate report, recall_delta means baseline recall minus slice recall, so a positive value is a regression. Failure summaries contain only counts and SHA-256 fixture hashes. They never serialize fixture text, entity surfaces, or model inputs. The suite remains offline when a predictor is provided and returns a structured skip reason when optional weights are not configured.

Opt-in compatibility smoke tests¶

The real-checkpoint tests are skipped by default with the environment variable needed to enable each layout:

OPENMED_INDIC_NER_COMPAT_LOCAL_MODEL=/srv/models/indic-ner \
  pytest tests/integration/test_indic_ner_checkpoint_compatibility.py::test_user_supplied_local_checkpoint_contract -q

OPENMED_INDIC_NER_COMPAT_REMOTE_MODEL=your-org/approved-indic-ner \
  pytest tests/integration/test_indic_ner_checkpoint_compatibility.py::test_explicit_remote_checkpoint_contract -q

An inaccessible configured checkpoint skips with an aggregate reason. A checkpoint that loads but violates the label, offset, logits, or privacy contract fails the test.