# Tiny Token Evaluation Demo `tiny_token_eval.jsonl` is a tiny fake evaluation set for smoke-testing the token-level baseline harness. It is intentionally small and not a benchmark. Each row has: - `prompt`: the question/context. - `answer`: the generated answer to score. - `hallucinated_spans`: answer substrings that should be labeled hallucinated. - `metadata`: notes explaining the toy case. The DGX evaluator converts `hallucinated_spans` into tokenizer-aligned `0/1` labels at runtime. That avoids hard-coding labels that only work for one tokenizer. Run it with: ```bash python scripts/eval_tiny_baselines_dgx.py \ --model Qwen/Qwen2.5-0.5B-Instruct \ --detector epr \ --detector lookback \ --device cuda \ --dtype bfloat16 ``` The wrapper writes token-level predictions to: ```text outputs/tiny_token_eval_predictions.jsonl ``` Use these numbers only as an end-to-end sanity check. Formal detector quality still needs a real labeled dataset.