LatentMD

Prompts, raw responses of 16 language models, and per-record scores of LatentMD, a benchmark for latent fence-boundary failures in LLM-generated Markdown: outputs that have the requested document structure but close a fenced code block where the model did not intend to, so that a CommonMark parser, a renderer or a code extractor reads a different document from the one the model meant. The accompanying code repository holds the scorer, the prompt generator, the inference runners and the analysis scripts.

What LatentMD measures

Every response is scored on two axes.

• Boundary (F2 ∧ F3): every opened fence is closed under the nesting the fences declare (F2), and the outer fence is longer than any inner same-family fence run (F3, undefined when there is no outer wrapper). The error types are E0 correct, E1 premature close, E2 unclosed fence and E3 latent delimiter collision.
• Content (C1 ∧ C2 ∧ C3): at least two headings, the requested code language, and at least the expected number of literal-syntax fenced blocks.

A response that passes the content checks and fails the boundary checks is a latent failure. F1 records instruction compliance under the A1 (no wrapper) and A2 (forced wrapper) conditions.

Layout

LatentMD-dataset/
├── prompts/                       the prompt suite (4,179 prompts, YAML)
│   ├── main_prompts.yaml              2,160  main A×B grid
│   ├── delimiter_family_prompts.yaml    288  D-axis: tilde vs backtick wrapper
│   ├── delimiter_length_prompts.yaml    432  D-axis: outer fence length
│   ├── delimiter_inner_prompts.yaml     432  D-axis: stated inner run
│   ├── cross_family_prompts.yaml        288  D-axis: cross-family nesting
│   ├── hint_prompts.yaml                432  prompt-hint ablation
│   ├── inline_prompts.yaml               27  inline code spans
│   ├── crossformat_prompts.yaml          90  Python and JSON probes
│   ├── natural_prompts.yaml              30  natural-prompt validation set
│   ├── stability_subset_prompts.yaml     30  stability subset (not counted in the 4,179)
│   ├── manifest.yaml                        split metadata
│   └── natural2/                            248 real user requests (see its README)
├── responses/                     raw responses, one JSONL per model (see responses/README.md)
│   ├── main_grid/<model>.jsonl        16 models × 2,160 prompts
│   ├── ablation/<model>.jsonl         16 models × 2,019 prompts
│   ├── natural2/<model>.jsonl         16 models × 248 real user requests
│   └── excluded_model/                Gemini-2.5-Flash-Lite (not in the cohort)
├── scores/                        per-record scores written by the released scorer
│   ├── main/, ablation/, natural2/    F1–F3, C1–C3, E0–E3 per response
│   ├── variants/                      F3 scoring-rule variants (sensitivity analysis)
│   ├── compliance/                    did the response implement the instructed configuration
│   └── rewrite/                       post-hoc safe-length rewrite baseline
└── consumer_panel/                four renderers and three extractors applied to every valid main-grid response

Schemas

Prompt records (prompts/*.yaml)

Field	Content
prompt_id	unique id, e.g. main_A1_B1_python_Python_35
category	main, d_family, d_length, d_inner_run, d_cross_family, hint, inline, crossformat, natural
prompt_text	the prompt the model receives
axis_conditions	per-axis levels, e.g. {"A": "A2", "B": "B3"}
slots	filled slot values (LANG, TASK, TASK_ID, TASK_LEVEL)

The A axis is the outer-wrapper instruction (A1 forbids a wrapper, A2 requires one, A3 says nothing), the B axis the inner nesting load (B1 one code block, B2 and B3 literal Markdown source with one or two inner blocks, B4 several Markdown elements), and the D, H and other axes the ablations named by the file.

Response records (responses/*/<model>.jsonl)

prompt_id, model, prompt_text (absent for natural2), response, finish_reason, latency, timestamp, metadata and usage; details in responses/README.md.

Score records (scores/*/<model>.jsonl)

Field	Content
prompt_id, model, a_condition, b_condition, lang, task_id	keys
boundary_correct, content_correct, cell	the two axes and the four-cell label (both_correct, latent_failure, format_only, both_wrong)
format_metrics	f1, f2, f3 (null when undefined), f3_variant, inner-run compliance fields
content_metrics	c1, c2, c3, counts of headings, code blocks and literal-syntax blocks, the matched language
errors	one-hot e0–e3
evasive	zero-width-space, reformatting and dropping flags

Only responses with finish_reason == "stop" are scored; the others are listed in the paper's truncation appendix.

Loading

from datasets import load_dataset
scores = load_dataset("json", data_files="scores/main/*.jsonl")
responses = load_dataset("json", data_files="responses/main_grid/gpt_4o.jsonl")

The dataset viewer shows the score, rewrite and consumer-panel files (one config each, a single test split: LatentMD is an evaluation set and has no training portion). The response files and the compliance scores are not previewed, because their fields differ across models and prompt categories; load them per file as above. The YAML prompt files load with yaml.safe_load. To score your own model, write records in the response schema and run the scorer of the code repository (python scripts/run_evaluation.py --responses <file-or-directory>).

Licenses and provenance

• Prompts and responses: CC BY-SA 4.0. The TASK slot text is taken from the instruction fields of McEval, which is released under CC BY-SA 4.0, and the suite inherits that licence with attribution. Model responses are the output of the listed models under their providers' terms of use.
• Real user requests (prompts/natural2/): the 116 WildChat-1M-derived prompts are released verbatim under ODC-BY with attribution; the 132 LMSYS-Chat-1M-derived prompts are released as conversation identifiers with a reconstruction script, because the LMSYS-Chat-1M license forbids redistributing the conversation text. The inclusion rubric and the per-candidate screening verdicts are in the same folder. One sensitive request is released as an identifier only and its responses are withheld (see responses/README.md).
• Both chat logs were de-identified by their authors before release, and the mining kept only conversations that the providers had not flagged for personal data or toxicity.

Anonymous authors, under review.