Repository navigation
Releases
lxask 1.1.0
What's new
- Precision adapter (
precision-round4.gguf): a second 37 MB LoRA routed per request. Restriction/proxy questions go to the restriction adapter; every other request goes to the precision adapter, trained on tasks that punish near-misses (wrong filenames, missing quotes, wrong flags, files touched that should be left alone). - Routing is automatic — a missing adapter just means the base model answers.
LXASK_PRECISION=offdisables it;--precision/--no-adaptercover batch mode.
Measured (held-out sets, commands executed in a sandbox)
- Strict execution success on generated file/text tasks: 35.0% -> 73.8%
- NL2SH-ALFA public benchmark: 62% vs 54% for the untuned model
- Everyday questions, right program: 96-99%
Binaries link glibc 2.28+: pick lxask-x86_64-v3 for x86 PCs since ~2015, lxask-x86_64 for older, lxask-aarch64 for ARM64. install.sh installs everything, or --from a directory.
lxask 1.0.0
lxask 1.0.0
Ask in plain English, get a Linux command checked against the man pages. It runs on your
machine: no network, account or API key after installing. One 1.1 GB install, about 1.3 GB
of memory while answering.
What 1.0 means
The command-line options, the settings file (~/.local/share/lxask/settings), the files
the installer puts on disk and the answer format are now stable. Changes that would break
them wait for 2.0.
Highlights
- Explains every part of the command, splits multi-step requests, keeps restrictions
("without lsof", "don't use sudo"), and asks back when a request is vague. - Checks every program and option against the manual and flags what it can't confirm;
warns before anything that deletes, overwrites or formats. It never runs a command for you. - Ctrl+G in bash and zsh (
lxask --shell-init), a server mode that keeps the model loaded
(answers in about 2 s),--command-onlyfor scripts,--bad/--sharefor reports. - Linux x86_64 (with a v3 build) and arm64. The model files are also on Hugging Face.
Measured
On the public NL2SH-ALFA
test set (300 shell tasks, each answer run in Docker by the InterCode-ALFA harness):
| solved | 95% CI | |
|---|---|---|
| lxask | 187/300 (62%) | 57-68% |
| the untuned Qwen2.5-Coder-1.5B it was trained from | 92/300 (31%) | 26-36% |
On the 237 tasks with no overlap with its training sources: 61% against 34%
(McNemar p < 10⁻¹³). Retries and retrieval were measured to make no difference on this
set; the score is the fine-tune's.
Where it misses, and every measurement with intervals and caveats:
docs/accuracy.md, docs/examples.md.
Install
curl -fsSLO https://github.com/arxhr007/lxask/releases/latest/download/install.sh
sh install.sh
Checksums are in SHA256SUMS. Offline install and upgrade: docs/install.md.
Licence
Code MIT. Model: a fine-tune of Qwen2.5-Coder-1.5B-Instruct (Apache 2.0). Training data
and documentation index licences: NOTICE.