DSLing2026/Tibetan-FormatEval

Tibetan FormatEval Benchmark — public 1,418-sample snapshot

This directory contains four aligned JSONL evaluation formats. Each file has 1,418 records in the same order and uses the same id values.

The public records retain dataset, subject, coarse domain, question, options, answer indices, gold answer text, and open-answer aliases. Internal source lineage and sample-split fields are intentionally omitted.

Files:

• main.jsonl: combined 4-way and 10-way representation.
• mcqa_4way.jsonl: four-choice multiple-choice evaluation format.
• mcqa_10way.jsonl: ten-choice multiple-choice evaluation format.
• open_short_answer.jsonl: open-answer evaluation format.

Evaluation scripts

The scripts/ directory contains standard-library-only Python 3.10+ tools for running an OpenAI-compatible model endpoint, scoring one model's predictions, and summarizing format-validity results across models.

Generate predictions

run_openai_compatible_eval.py accepts either the 4-way or 10-way JSONL file with --task mcqa, or the open-answer file with --task open:

python scripts/run_openai_compatible_eval.py \
  --gold mcqa_4way.jsonl \
  --task mcqa \
  --model YOUR_MODEL_NAME \
  --base-url https://YOUR_ENDPOINT/v1 \
  --api-key "$API_KEY" \
  --output predictions/YOUR_MODEL_NAME/mcqa_4way.jsonl \
  --concurrency 1

The runner writes resumable prediction JSONL records with the schema {"id": "...", "prediction": "..."} plus model and request metadata. Use an environment variable or a private --shell-file for credentials; do not commit keys to the repository.

Score predictions

evaluate_predictions.py supports letter/index MCQA predictions and normalized open short answers. It writes both a summary JSON and a matching .per_item.jsonl file:

python scripts/evaluate_predictions.py \
  --gold mcqa_4way.jsonl \
  --predictions predictions/YOUR_MODEL_NAME/mcqa_4way.jsonl \
  --task mcqa4 \
  --output results/YOUR_MODEL_NAME/mcqa_4way.json

python scripts/evaluate_predictions.py \
  --gold mcqa_10way.jsonl \
  --predictions predictions/YOUR_MODEL_NAME/mcqa_10way.jsonl \
  --task mcqa10 \
  --output results/YOUR_MODEL_NAME/mcqa_10way.json

python scripts/evaluate_predictions.py \
  --gold open_short_answer.jsonl \
  --predictions predictions/YOUR_MODEL_NAME/open_short_answer.jsonl \
  --task open \
  --output results/YOUR_MODEL_NAME/open_short_answer.json

Compare multiple models

After each model has the three scored summaries and per-item files under its own result directory, run:

python scripts/analyze_format_validity.py \
  --result-root results \
  --output results/format_validity_summary.json