AdapterLab Support Intent Tiny

A real, CPU-trained LoRA adaptation of BERT Tiny for six English support intents: account, billing, cancellation, delivery, login, and technical. Built by Inioluwa Atanda as the reference experiment for AdapterLab, a training product with dataset validation, durable job execution, evaluation, export and inference.

Try the browser demo · Inspect the dataset

This release demonstrates an inspectable training and deployment pipeline. It is an experimental classifier with modest performance, not a production-ready model.

Measured results

The grouped, synthetic holdout contains 30 examples, five per label.

Training time is one workspace CPU run with two PyTorch threads on x86_64; it is not a general latency benchmark. The exported classifier has 4,386,694 parameters after merging; the wrapper count includes LoRA and classifier wrapper copies.

Billing and cancellation each have zero test F1. Do not infer strong general support accuracy from the aggregate score. The baseline in evaluation.json is the pretrained encoder with an untrained classification head, not a trained head-only competitor. No state-of-the-art or multilingual capability is claimed.

Training and provenance

• Base: google/bert_uncased_L-2_H-128_A-2, revision 30b0a37ccaaa32f332884b96992754e246e48c5f.
• LoRA rank 8, alpha 16, dropout 0.05, query/value targets; classifier trained.
• AdamW, learning rate 0.001, weight decay 0.01, batch 8, four epochs, seed 42.
• Maximum input length 96 tokens; best checkpoint chosen by validation macro F1, with validation loss as the tie-breaker.
• 180 train / 30 validation / 30 test examples; template groups are disjoint.
• Actual training date: 27 October 2026, recorded in run.json.

provenance.json pins the original source revision, run ID and package versions. The original evaluation, per-class results, confusion matrix, training history, dataset fingerprint and checksums are included. This publication reused the existing run; it did not retrain or change its measurements.

Load the merged classifier

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Ineeoluwa/adapterlab-support-intent-tiny",
    trust_remote_code=False,
    device=-1,
)
print(classifier(
    "I cannot sign in after resetting my password.",
    top_k=6, truncation=True, max_length=96,
))

The root contains merged model.safetensors and its tokenizer. The original adapter and tokenizer are retained under adapter/. To load that adapter, create the pinned base with num_labels=6 and the id2label / label2id mappings in config.json, then use PeftModel.from_pretrained(base, repo, subfolder="adapter"). Creating a generic two-label base will not load the six-label head correctly.

onnx/model.onnx is float32, opset 17, with dynamic batch and sequence dimensions. The browser demo uses Transformers.js and a local ONNX Runtime WASM worker. PyTorch-to-ONNX prediction agreement is 30/30 on the test set; maximum logit error is below 1e-5. Singleton, mixed-length, truncated and Unicode batches were also compared. This is export equivalence, not an independent accuracy study.

Intended use and limits

Useful for learning PEFT, checking grouped splits, reproducing a small experiment, and inspecting browser inference. The examples are templated synthetic English, not customer tickets. Reference tokens, template choice and the tiny holdout limit representativeness. The model has no rejection class, calibration, robustness, fairness, real-world or Yoruba evaluation. Softmax scores should not be treated as calibrated probabilities. Avoid sensitive or consequential use.

License

Apache-2.0, including the base model's license. Dataset examples are released separately under CC0-1.0. Original base model credit: Google / the BERT authors.