Try
↑↓Move↵OpenCtrl↵New tabescClose

Datasets

I published these four datasets on Hugging Face under open licenses and trained the models below on them. Two are training data for coding-agent guardrails, one for shell commands and one for JavaScript and TypeScript code review. The other two are Laravel 11 question and answer sets, now marked legacy.

Datasets
4
Rows
151,025
Model releases built on them
12
Dataset downloads, all time
2,243

All datasets

JS/TS Code Review Decisions

Benchmark

Four small models were fine-tuned on the same 9,600 snippets and scored on 15,600 test decisions. The list also has a TF-IDF baseline and two reference rows. Pick a metric to re-rank it.

Higher is better. The mean F1 over the 10 flaw questions at a 0.5 cutoff. It is the primary metric.

  1. d1-3B on Hugging Face (opens in a new tab)3.1B · LFM2.5-VL-3B with Liquid AI’s d1 decision readout0.917
  2. d1-omni-600M on Hugging Face (opens in a new tab)587M · LFM2.5 encoder with a decision head0.771
  3. GLiNER2.5-Decide on Hugging Face (opens in a new tab)486M · GLiNER2 DeBERTa-v3 encoder with a classification head0.687
  4. TF-IDF + logistic regressionBaseline fitted on the train split0.502
  5. NanoJev on Hugging Face (opens in a new tab)596M · Qwen3-0.6B with a choice head0.452
  6. UniformGives every option the same probability0.237
  7. PriorPredicts the train label frequencies and ignores the input0.000
  • Fine-tuned on train
  • Baseline
  • Reference

The full benchmark

CLI Command Risk

Rows by risk level

  1. Safe39.9%49,890 rows
  2. Low8.3%10,361 rows
  3. Moderate8.3%10,364 rows
  4. High18.8%23,539 rows
  5. Critical24.6%30,769 rows

All levels and categories

Models

huggingface.co/yannelli (opens in a new tab)

Trained on CLI Command Risk

Trained on JS/TS Code Review Decisions

Trained on Laravel 11 QA and Long Form