steph-speak-train

Supervised fine-tuning data behind the steph-speak models: terminal command outputs (or short session histories with a question) paired with the single spoken line a blind or visually impaired user should hear. It is the exact training mix of steph-speak-granite-4.0-h-350M-v2-GGUF; the Gemma model was trained on a closely related mix of the same sources.

Format

One JSON object per line:

Field	Meaning
source	neutral category of the example (never a path, host or person)
messages	system, user, assistant turns in the exact format the models were trained on

The user turn holds the command, its status, the digested output and, for recent versions, a short block of precomputed facts. The assistant turn is the label: at most 15 words, names and numbers copied from the output, nothing invented.

Content

Part	Description
Summaries	spoken summaries of real command outputs captured on Linux and macOS machines while developing steph, anonymized
Feature prompts	detail, progress and stream prompts built on the same outputs
History questions	questions about a short session history, including questions whose answer is absent and must be declined
Language fix	examples with an English prompt over French command output, answered in English

English and French. Training split: 4,571 examples (French prompt 3,098, English prompt 1,473); validation split: 221 (168, 53). This is the exact mix the small model steph-speak-granite-4.0-h-350M-v2 was trained on.

How the labels were made

Most labels come from Qwen3.5 27B; the remaining feature, question and language-fix rows come from Qwen3.5 9B (both Apache 2.0, no restriction on using outputs for training), at temperature 0. A second pass by a Qwen3.5 9B judge removed answers containing invented facts. Rows that no longer ground against their input were dropped. These are model labels, not human labels.

Anonymization and checks

User, host and project names were replaced, IP addresses kept in private (RFC 1918) or documentation ranges, MAC addresses, UUIDs and serial numbers removed. An automated scanner with a private list of names runs on the published files. Rows that match a fingerprint of the held-out benchmarks, or that carry the benchmark canary string, are removed from the mix.

Intended and out-of-scope use

Intended: training and studying small models that read terminal output aloud. Not intended: any task unrelated to terminal output. Do not use steph-speak-bench test data for training.

Limitations

• Captured on the machines of one developer: tool coverage follows that developer's work (compilers, git, package managers, network, disks, containers); no Windows.
• Labels are produced by a model and may contain residual errors.
• English and French only.

License

CC BY 4.0. Cite as:

@misc{steph_speak_train_2026,
  author = {Parra, Luis},
  title  = {steph-speak-train},
  year   = {2026},
  url    = {https://huggingface.co/datasets/lsmdx/steph-speak-train}
}