# Benchmarks and review files

This directory includes:

## 500-sample mixed-direction review set
Source-only benchmark for human review:
- `kashmiri_benchmark_500_source_only.jsonl`
- `kashmiri_benchmark_500_stage1_safe_pred.jsonl`
- `kashmiri_benchmark_500_stage1_safe_pred.summary.json`
- `kashmiri_benchmark_500_stage1_safe_pred.xlsx`
- `kashmiri_benchmark_500_stage1_safe_pred.summary.xlsx`

## 200-sample ks→en evaluated benchmark
- `benchmark_simple_ks2en_200.jsonl`
- `benchmark_simple_ks2en_200.stage1_eval.json`

The 500-sample file is for manual review. The 200-sample file includes references and saved evaluation output.