umt5-small-mic

Slim model weights, executable usage and framework expectations in one repository.

• Contract manifest
• Transformers usage
• Transformers expectations
• Fixture provenance

Contract

MIC: T5 text-to-text generation

Model integration contract (MIC) for one usage path, across the checkpoints that share it. This repository holds the contract code and the manifest; each framework keeps its own expected results and pins this repository by revision. The model repositories are unchanged. It is a test artifact, not a model: exclude it from model search.

Usage path: AutoTokenizer + AutoModelForSeq2SeqLM.generate, batched.

File	What it is
integration.json	Manifest: checkpoints, which fixture covers each, frameworks, entry points, variants, unsupported targets, comparison rules
integrations/transformers/usage.md	Transformers contract, written as doc-builder runnable markdown: the same blocks are documentation and test

Checkpoint	Slim fixture (every framework PR)
google-t5/t5-small	hf-internal-testing/mic-t5-small-slim
google/umt5-small	hf-internal-testing/mic-umt5-small-slim (real checkpoint unsupported: config.json has no model_type, so AutoModelForSeq2SeqLM cannot load it; the fixture (saved with model_type umt5) covers the UMT5 code path)

Checked outputs: generated_tokens, generated_text, label_logprobs, decoded_roundtrip, training.loss, training.gradients, training.updated.

A fixture pass and a real-checkpoint pass have different scopes; either establishes agreement on fixed inputs and a recorded runtime, not correctness on arbitrary inputs.

Slim model

Slim fixture: proto-umt5-small-slim

A 16.6M-parameter fixture for the T5 MIC, covering google/umt5-small. It is a test artifact, not a usable model, tagged test-artifact to keep it out of candidate-model search.

• Fixture key d2f1e2c7f664. It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
• Same architecture, smaller config: UMT5ForConditionalGeneration, shrunk by the generic rule in sweep.py (num_layers 2, num_decoder_layers 2).
• Processor files copied byte for byte from google/umt5-small: generation_config.json, special_tokens_map.json, spiece.model, tokenizer.json, tokenizer_config.json.
• Random weights (seed 17). They passed the input-sensitivity gate: every perturbation of the contract's inputs moved the logits by at least 0.001 (weakest: attention_mask:all_ones, 3.09), so no training was needed.
• Same tensor names as the real checkpoint. The real checkpoint has every parameter, and so does the fixture.
• Finite in float32, bfloat16, and float16 (forward, loss, gradients).

Built by build_fixtures.py; build.json records the pinned sources, gate results, and file hashes. This repository includes the slim weights, usage and framework expectations; rebuild the other files with python example/build_fixtures.py --fixture proto-umt5-small-slim.