---
language:
- en
- th
- vi
- zh
pretty_name: HSClassify Benchmark
size_categories:
- n<1K
task_categories:
- text-classification
tags:
- hs-codes
- trade
- customs
- harmonized-system
- benchmark
dataset_info:
  features:
  - name: text
    dtype: string
  - name: expected_hs_code
    dtype: string
  - name: category
    dtype: string
  - name: language
    dtype: string
  - name: notes
    dtype: string
  config_name: default
  splits:
  - name: test
    num_examples: 78
configs:
- config_name: default
  data_files:
  - split: test
    path: benchmark_cases.csv
---

# HSClassify Benchmark

Benchmark suite for evaluating the [HSClassify](https://github.com/JamesEBall/HSClassify_micro) HS code classifier.

## Results (latest)

| Metric | All Cases | In-Label-Space |
|--------|-----------|----------------|
| Top-1 Accuracy | 79.5% | 88.6% |
| Top-3 Accuracy | 82.0% | 91.4% |
| Top-5 Accuracy | 83.3% | 92.9% |
| Chapter Accuracy | 89.7% | 95.7% |

### By Category

| Category | N | Top-1 | Top-3 | Top-5 |
|----------|---|-------|-------|-------|
| easy | 27 | 96.3% | 100% | 100% |
| edge_case | 21 | 71.4% | 76.2% | 81.0% |
| multilingual | 20 | 100% | 100% | 100% |
| known_failure | 10 | 10.0% | 10.0% | 10.0% |

## Test Cases

78 hand-crafted cases in `benchmark_cases.csv` across four categories:

- **easy** (27): Common goods the model should classify correctly
- **edge_case** (21): Ambiguous queries, short text, brand names
- **multilingual** (20): Thai, Vietnamese, and Chinese queries
- **known_failure** (10): Documents current blind spots and label-space gaps

## Usage

Requires a trained [HSClassify_micro](https://github.com/JamesEBall/HSClassify_micro) model directory as a sibling folder (or pass `--model-dir`).

```bash
# Basic benchmark (~10s)
python benchmark.py

# Custom output path
python benchmark.py --output results/out.json

# With per-class split analysis
python benchmark.py --split-analysis

# Point to model directory explicitly
python benchmark.py --model-dir /path/to/HSClassify_micro
```

## Split Analysis (training data)

Replicates the 80/20 stratified split from model training to report:
- Worst 15 HS codes by F1 score
- Top 20 cross-chapter confusions
- Overall accuracy: **77.2%** (matches training baseline)