YuvrajSingh9886/asciitermdraw-bench-public
ASCIITermDraw-Bench — Public Examples 12 public example tasks from ASCIITermDraw-Bench, a benchmark for evaluating whether language models can generate and edit structured ASCII diagrams. The full benchmark has 80 private, held-out tasks used for actual scoring — those are not distributed here. This dataset is a separate, hand-authored set of 12 tasks (one easy, one medium, one hard per category) in the exact same format, so anyone can see what a task looks like and run the… See the full description on the dataset page: https://huggingface.co/datasets/YuvrajSingh9886/asciitermdraw-bench-public.
ASCIITermDraw-Bench — Public Examples
12 public example tasks from ASCIITermDraw-Bench, a benchmark for evaluating whether language models can generate and edit structured ASCII diagrams.
The full benchmark has 80 private, held-out tasks used for actual scoring — those are not distributed here. This dataset is a separate, hand-authored set of 12 tasks (one easy, one medium, one hard per category) in the exact same format, so anyone can see what a task looks like and run the scoring pipeline without access to the private set.
Repository: https://github.com/YuvrajSingh-mist/ASCIITermDraw-Benchmark
Categories
box-layout-basics— box drawing and ASCII layout basicsnetwork-topology-diagrams— network, systems, and cluster topology diagramsdiagram-editing— image-conditioned edit tasks (source diagram in, edited diagram out)software-architecture-diagrams— canonical software architecture sketches
Layout
<category>/<difficulty>/<task_id>/
prompt.txt instruction sent to the model under test
assertions.json ground-truth structural requirements (labels, entity
count, edges, edge labels)
reference.ascii gold-standard ASCII diagram
reference.png reference.ascii rendered to an image
source.ascii/.png diagram-editing tasks only: the "before" diagram
vlm_judge_prompt.txt exact prompt template sent to the LLM judgeTask ids 0.1–0.12, matching the same <category>/<difficulty>/<task_id> layout the benchmark's own tooling expects, so this set is directly usable as a --tasks target for the repository's generation/judging scripts:
uv run run-model --model Qwen/Qwen3.7-Plus --tasks public_dataset --outputs outputs/public_demo
uv run judge-geval --provider openai --model gpt-5.4 --tasks public_dataset --outputs outputs/public_demo/qwen3.7-plus --results outputs/public_demo/results.csvThese 12 tasks are not part of the benchmark's official scoring — they're documentation/demonstration data.
License
- Dataset / task content: CC BY 4.0 — free to share and adapt, including commercially, with attribution to Yuvraj Singh (name + link; indicate changes if you modify). See `LICENSE-DATASET`.
- Code / tooling (runners, renderers, judge): Apache-2.0.
The private 80-task evaluation set is not distributed and is not covered by this CC BY grant.
Citation
If you use ASCIITermDraw-Bench — the benchmark, tasks, tooling, leaderboard, or results — please credit this work and cite it as:
@misc{singh2026asciitermdrawbench,
title={ASCIITermDraw-Bench: Benchmarking ASCII Diagram Generation and Editing},
author={Yuvraj Singh},
year={2026},
howpublished={\url{https://github.com/YuvrajSingh-mist/ASCIITermDraw-Benchmark}},
}<!-- ad-footer --> ---
Released freely — support more experiments like it:
 
More: <https://huggingface.co/YuvrajSingh9886>
