--- language: - en tags: - video-generation - agent-evaluation - benchmark - execution-traces size_categories: - n<1K pretty_name: VAHBench — Five Briefs and Curated Examples configs: - config_name: prompts data_files: - split: test path: benchmark/prompts.jsonl --- # VAHBench Five short-drama test prompts and a compact showcase aligned with [the VAHBench GitHub repository](https://github.com/TokenRhythm/VAHBench). This repository contains **5 prompts, 15 example runs (3 per prompt), and 66 final episode videos**, together with their execution traces. ## 1. Test prompts and requirements — `benchmark/` | Brief | Full prompt | Episodes × duration | Aspect ratio | Requirements | |---|---|---|---|---| | B1 | [The Woman Behind the Numbers](benchmark/prompts/B1_the_woman_behind_the_numbers.md) | 3 × 60 s | 9:16 | [JSON](benchmark/requirements/B1_the_woman_behind_the_numbers.json) | | B2 | [Terms of the Borrowed Vow](benchmark/prompts/B2_terms_of_the_borrowed_vow.md) | 3 × 120 s | 9:16 | [JSON](benchmark/requirements/B2_terms_of_the_borrowed_vow.json) | | B3 | [The Algorithm Knows My Name](benchmark/prompts/B3_the_algorithm_knows_my_name.md) | 5 × 120 s | 9:16 | [JSON](benchmark/requirements/B3_the_algorithm_knows_my_name.json) | | B4 | [The Nineteenth Floor](benchmark/prompts/B4_the_nineteenth_floor.md) | 5 × 240 s | 16:9 | [JSON](benchmark/requirements/B4_the_nineteenth_floor.json) | | B5 | [The Last Clause](benchmark/prompts/B5_the_last_clause.md) | 6 × 300 s | 9:16 | [JSON](benchmark/requirements/B5_the_last_clause.json) | - [Prompt and requirement index](benchmark/index.json): B1–B5 identifiers, prompt hashes, source references, and structured requirements. - [Five-row JSONL table](benchmark/prompts.jsonl): the `prompts` dataset configuration / `test` split, containing full prompts and core specifications. - [Benchmark configuration](benchmark/config/benchmark.json): dependencies, numbering, scope, and interpretation notes. - [Evaluation defaults](benchmark/config/default.yaml) and [scoring thresholds](benchmark/config/thresholds.yaml): byte-identical snapshots from GitHub commit `36c00ec`. - [Evaluator Python dependencies](benchmark/config/evaluator-requirements.txt). The English prompts are copied verbatim from the initial user task in the archived traces and cross-checked between Hermes and OpenSquilla. Structured requirements include the original field text as well as parsed episode counts, duration, characters, dialogue constraints, timing bands and deliverables. The original prompt is authoritative, including its optional-supplement wording and final acceptance criteria. Requirements are not agent outputs. [`prompt.md`](prompt.md) is a generated compatibility copy containing the same five prompts. B1–B5 are the canonical IDs; `Prompt 1` through `Prompt 5` headings in this compatibility file preserve the GitHub evaluator’s input format. ## 2. Showcase data — `examples/` See the **[15-run gallery and all episode links](examples/README.md)**. Each brief has one Hermes, one OpenClaw and one OpenSquilla run, all with `deepseek-V4-flash`, keeping the harness/model combination aligned with the GitHub examples. The three existing GitHub gallery runs are preserved byte-for-byte under the B1–B5 directory names; twelve additional runs cover the remaining brief/harness pairs. ```text benchmark/ prompts/ # five original prompts requirements/ # per-prompt structured requirements config/ # benchmark and evaluator configuration index.json prompts.jsonl # Hugging Face prompts/test table examples/ /deepseek-V4-flash/B_/ ./trajectory-exports/trajectory/ index.json provenance.json media.json validation.json prompt.md # evaluator compatibility index ``` The showcase contains final episode files and original trajectory exports. Temporary renders, source clips, production assets and unselected runs remain in the **[complete pre-curation archive](https://huggingface.co/datasets/TokenRhythm/VAHBench/tree/raw-before-curation-2026-11-18)** at revision `f1ec38d2ff57297db54a4f86f1d5d3920a6aa231`. Current paths, archived source paths, sizes, hashes and the numbering history are recorded in [`examples/provenance.json`](examples/provenance.json). This is a showcase subset, not the full dataset used to report 240 runs. ### Use with the GitHub evaluator Download this dataset (authentication is required while the repository is private). Keep the root `prompt.md` beside the `examples/` directory. From the GitHub checkout, with its dependencies and `ffmpeg`/`ffprobe` installed: ```bash python run.py --data /path/to/VAHBench-dataset/examples --discover-only python run.py --data /path/to/VAHBench-dataset/examples --out-dir out/showcase --allow-partial ``` Use the settings from the GitHub checkout; the files under `benchmark/config/` are provenance snapshots. Configure the extraction-model credential through the environment as described in the GitHub README. No credentials are included. The bundled reference detectors produce **partial** evaluations, not exact reproductions of paper scores. This subset omits intermediate source clips, so source-dependent checks may lack evidence. Use the archive for full run artifacts. The three paper-reported example scores are attributed in the run index; the twelve additional examples have not been assigned scores here. ### Validation and provenance All 172 copied files were verified against the source repository's object hashes. All 31 files in GitHub's three example runs match exactly. The GitHub input-discovery code recognizes 15 runs and correctly matches all five briefs. All 66 videos pass first-frame and last-second decode checks; JSON and JSONL trace exports parse successfully. This is structural validation, not a full video decode or a new benchmark evaluation. Repository visibility is unchanged. No new media license is asserted by this curation; the GitHub code license and third-party model terms retain their scope.