ahmedsamirtarjama/shellsafe-bash-sft

shellsafe bash SFT

Natural-language request to Bash command pairs in chat format, used to train shellsafe-qwen2.5-coder-0.5b. That model is the weight average of one fine-tune on v4 and one on v5.

Config	Train	Validation
v4	72,055	1,246
v5 (default)	77,774	1,246

Both configs share the same base rows; they differ in the teacher-generated rows (round 1 in v4, round 3 in v5).

Format

One JSON object per line:

{"messages": [{"role": "system", "content": "You are a bash command generator. ..."},
              {"role": "user", "content": "Request: find all .tmp files not accessed in 7 days"},
              {"role": "assistant", "content": "find . -name '*.tmp' -atime +7"}],
 "id": "compose:1260:7", "source": "compose", "mode": "none"}

Sources

source says where a row came from:

source	What it is	Origin and license
tldr*	requests written by Qwen3-Coder-30B-A3B-Instruct for tldr-pages example commands	commands from tldr-pages contributors, CC BY 4.0
compose*	multi-step requests and commands built from tldr examples by the same model	derived from tldr-pages, CC BY 4.0
nl2sh*	the train split of NL2SH-ALFA	MIT
targeted:*	requests about a fixture file system written by Qwen3-Coder-30B-A3B-Instruct, kept only when two independent teacher answers printed the same output in a sandbox	generated for this dataset

Candidate pairs were run in a sandbox and judged by the local teacher before they were kept. Suffixes: +fix is a command the teacher-judge corrected (the correction also had to run); +rl is a row that got a second opinion from API models (Kimi-K3, DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next), whose command was replaced only when two of them called the original wrong and the replacement ran in the sandbox.

Because of the tldr-derived rows the dataset as a whole is released under CC BY 4.0. Please credit tldr-pages and the NL2SH-ALFA authors when you use it.

Decontamination

No NL2SH-ALFA test items or InterCode-Corrections data are included. No training request matches an NL2SH-ALFA test request exactly, and teacher-generated requests with word overlap (Jaccard) of 0.6 or more with any test request were dropped.

Notes

• Emails, keys and host names in the data are made-up placeholders (for example john.doe@gmail.com).
• Some commands are destructive (for example rm -rf); the data teaches what a command looks like, not whether it is wise to run it.