# Build and Evaluate Adaptive Text Classifiers

Small labeled datasets and changing categories make text classification difficult to maintain. This skill guides incremental classifier updates and held-out evaluation with the adaptive-classifier library.

## Install

```bash
npx skillstore add lambdasec/adaptive-classifier
```

## Metadata

- Status: approved
- Slug: lambdasec-adaptive-classifier
- Skillstore revision: r1
- Version status: missing
- Tree hash: 2e64ffb79e3b9908530fdd4313bc5dabbf57bc1d50161a2c34b753589a058b5d
- Author: lambdasec
- GitHub username: lambdasec
- License: Apache-2.0
- Repository: https://github.com/lambdasec/skills/tree/7d791206069498962a85a0f1ba95f7468dcf82f7/skills/adaptive-classifier
- Ref: 87c9669baf85385959f9b0a344dc9dfc8d978f9c
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 75
- Public page: https://skillstore.pages.dev/skills/lambdasec-adaptive-classifier
- Manifest: https://skillstore.pages.dev/api/skills/lambdasec-adaptive-classifier/manifest

## Capabilities

- Guides installation checks and API signature inspection for adaptive-classifier.
- Provides examples for single-text, batch, and multilabel predictions.
- Explains incremental example updates, new classes, and class statistics.
- Defines held-out evaluation with baselines, per-class metrics, calibration checks, and order-sensitivity testing.
- Documents local save and load workflows, ONNX options, and download approval requirements.
- Explains strategic prediction modes and robustness evaluation for adversarial text.

## Use Cases

- Route changing support intents: Build a ticket classifier from labeled examples, add emerging intents, and reevaluate existing categories after updates.
- Evaluate small classification datasets: Compare an adaptive classifier against simple baselines using held-out data, per-class metrics, and learning curves.
- Assess moderation triage robustness: Explore multilabel predictions and strategic robustness for moderation queues while retaining human review for consequential decisions.

## Prompt Templates

### Check whether the library fits

```
Assess whether adaptive-classifier fits my ticket-routing task with 20 labeled examples per class. Identify missing information and ask before downloading anything.
```

### Build a first classifier

```
Use my labeled texts to draft a local intent classifier with held-out evaluation. Check the installed version and request approval before downloading weights.
```

### Add a new category

```
Add my new billing examples to an existing classifier. Check label consistency and compare held-out performance for old and new classes.
```

### Compare configurations and robustness

```
Design validation-only comparisons for pooling, blending weights, ONNX inference, and strategic robustness. Include order sensitivity, calibration, and a final untouched test evaluation.
```

## Limitations

- Requires the adaptive-classifier library; the skill does not install it automatically.
- Documents version 0.2.0; model-dependent examples were not executed by the skill author.
- Model weights may require approved downloads, and classification quality depends on the encoder and labeled examples.
- Updates can change earlier predictions; scores are not guaranteed probabilities, and high-stakes decisions require human review.

## Best Practices

- Separate training, validation, and test data before adding examples, and remove near-duplicates across splits.
- Request approval before model downloads, keep remote model code disabled, and publish only when explicitly requested.
- Reevaluate all classes after updates and report per-class metrics with sample sizes.

## Anti Patterns

- Evaluating on training examples or tuning settings against the final test split.
- Treating blended prediction scores as calibrated probabilities or upstream benchmark figures as guaranteed outcomes.
- Automating high-stakes decisions without human review or ignoring regressions after new labels.

## Security Audit

- Audited at: 2026-10-28T06:31:15.63\+00:00
- Summary: All 32 static findings are false positives: 31 match Markdown backticks, and one matches a documentation URL. No evidence found of prompt injection, credential access, or data exfiltration in the reviewed files. Model downloads require approval, remote model code stays disabled, and model-dependent examples were not executed during this audit.

## Stats

- Views: 0
- Downloads: 0
- Favorites: 0
- Popularity score: 0