android/sample/scripts/TEST_DATA.md

Test data

Debug gate asset (committed)

app/src/debug/assets/gate_fixtures.json holds 152 captured official requests: all 144 rows of SemIf authored144.jsonl (MIT, GitHub TheoLeeCJ/SemIf at ca3ba65f), the model card's two examples and six invented requests. Each row carries the request (text, prompt, labels), the token IDs and <<LABEL>> positions the official gliclass 0.1.20 pipeline fed the model (CPU fp32, batch 1), the window, the fp32 logits and the official single-label and multi-label (0.5) results. All 152 fit the 128-token window, so the committed asset exercises only the s128 graph. The asset is written by the conversion run's scripts/app_gate_fixtures.py --subset.

The full gate also used 200 ag_news test rows and 200 banking77 test rows, 70 of which need the 256-token window. Their text is not redistributed (the ag_news dataset card gives its license as "unknown"; banking77 is CC BY 4.0). To run that gate, regenerate the asset from the conversion run without --subset, which writes all 552 rows to the same path. Do not commit that file; restore the subset with git checkout -- app/src/debug/assets/gate_fixtures.json.

Optional JVM reference data

The JVM tests compare the Kotlin host with the official gliclass 0.1.20 pipeline and the Python tokenizers library. The reference data is not distributed with the source. Without gliclass.fixtures, or with a missing directory or file, JUnit reports an assumption skip. Skips are not parity evidence.

./gradlew :app:testDebugUnitTest -Pgliclass.fixtures=/path/to/data
# Equivalently: -Dgliclass.fixtures=/path/to/data

Reports go to app/build/reports/parity/, never to the data directory.

data/
  fixtures/oracle.json             # 552 official pipeline calls (CPU fp32, batch 1): linearized
                                   # string, captured input_ids, label_positions, logits, softmax,
                                   # sigmoid, single-label and multi-label (0.5) results, fits
  fixtures/requests.json           # the 552 requests: text, prompt, labels
  fixtures/tokenizer_stress.json   # edge strings with Python tokenizers ids ([CLS]/[SEP] added)
  tokenizer.json                   # or host_assets/tokenizer.json, or the conversion run's
                                   #   src/gliclass-edge-v3.0/tokenizer.json
  tok_embeddings_fp16.bin          # or host_assets/, or the run's exports/tables/

Test	Needs	Checks
GliclassTokenizerTest	oracle, stress, tokenizer	ids of the 552 linearized oracle strings and of every stress string; special tokens and template
GliclassInputsTest	oracle, tokenizer, table	linearization, ids, <<LABEL>> positions, smallest window, padded ids / attention / routing / embeddings of all 552 requests; rejection rules; float16 upcast of every bit pattern
GliclassDecoderTest	oracle	oracle logits → the pipeline's single-label and multi-label results (labels exact, scores within 1e-6); tie, threshold and repeated-label rules
GateFixturesTest	oracle, tokenizer, table, debug asset	every row of the debug asset (152 committed, or 552 regenerated) equals the oracle; host inputs identical for every row; declared row and window counts; the prefilled request is the oracle's inv_02

The oracle, the requests and the stress strings are written by the conversion run's scripts/source_oracle.py, scripts/make_fixtures.py and scripts/tokenizer_stress.py.