Aaron Thomas commited on
Commit
2e118cd
·
verified ·
1 Parent(s): bd47a7a

Model card

Browse files
Files changed (1) hide show
  1. README.md +87 -0
README.md ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
4
+ library_name: gguf
5
+ pipeline_tag: text-generation
6
+ language:
7
+ - en
8
+ tags:
9
+ - linux
10
+ - bash
11
+ - shell
12
+ - command-line
13
+ - gguf
14
+ - llama.cpp
15
+ - lora
16
+ ---
17
+
18
+ # lxask: Linux commands from plain English, offline
19
+
20
+ `lxask` turns a plain-English request ("find files bigger than 100 MB in /var, newest first")
21
+ into a Linux command, with each part explained, and checks it against the man pages and
22
+ tldr before showing it. It runs on your machine: no network, account or API key after
23
+ install. The program is a single Rust binary; this repository holds the model files it uses.
24
+
25
+ - Code, install script and docs: https://github.com/arxhr007/lxask
26
+ - Docs site: https://arxhr007.github.io/lxask/
27
+ - Version of these files: lxask v0.11.0
28
+
29
+ ## Files
30
+
31
+ | file | size | what it is |
32
+ |---|---|---|
33
+ | `lxask-round4-Q4_K_M.gguf` | 986 MB | the chat model: Qwen2.5-Coder-1.5B-Instruct, fine-tuned, 4-bit (Q4_K_M) |
34
+ | `restrict-round4.gguf` | 37 MB | LoRA adapter, switched on only for requests that rule a program out ("without lsof") or mention a proxy |
35
+
36
+ sha256:
37
+
38
+ ```
39
+ 023ee206403c7fc96a9be13e1aec761dd798d73c6d2f876a6c09c392dbf9cbc5 lxask-round4-Q4_K_M.gguf
40
+ e77b3d86aab83aedf1a464993599c4b503d59d80b3fa7147b420221c214c8549 restrict-round4.gguf
41
+ ```
42
+
43
+ These are not a standalone chat model. The accuracy comes from lxask's pipeline around the
44
+ model: retrieval from man pages and tldr, a validator that checks every program and option,
45
+ retries, and fixed rules. Used raw, the model alone is weaker (see the benchmark below).
46
+
47
+ ## Use
48
+
49
+ Install lxask (it downloads these files):
50
+
51
+ ```
52
+ curl -fsSLO https://github.com/arxhr007/lxask/releases/latest/download/install.sh
53
+ sh install.sh
54
+ lxask "show the 5 biggest files in this folder"
55
+ ```
56
+
57
+ The model alone runs in llama.cpp (`llama-cli -m lxask-round4-Q4_K_M.gguf`), but without
58
+ the pipeline you lose the checks.
59
+
60
+ ## Training
61
+
62
+ Base: Qwen2.5-Coder-1.5B-Instruct (Apache 2.0). Two stages, then a small add-on:
63
+
64
+ 1. Supervised fine-tuning (LoRA, about 39,000 examples, one epoch).
65
+ 2. Reinforcement learning (GRPO) on about 1,550 file and text tasks, rewarded when the
66
+ command actually ran in a sandbox and did the job. Held-out execution success rose from
67
+ 52% to 78% (raw, before the pipeline's checks).
68
+ 3. The 37 MB adapter, trained on requests that rule out a program and on proxy settings.
69
+
70
+ Data: tldr-pages (CC BY 4.0), NL2Bash (GPL-3.0, with macOS-only commands removed), and
71
+ questions and answers written by language models, every command checked by a validator and
72
+ the answers reviewed by a stronger model. The model-written data is not published. Every
73
+ training round, including the ones that were dropped:
74
+ https://github.com/arxhr007/lxask/blob/main/docs/results.md
75
+
76
+ ## Limits
77
+
78
+ - 1.5B parameters. It misses by a detail (a flag, a field number) more often than a large
79
+ model would, which is why every answer is explained and anything unconfirmed is flagged.
80
+ - English requests; Linux only (macOS and BSD flags differ).
81
+ - It never runs commands for you. Read a command before you run it.
82
+ - Licence lineage: Qwen2.5-Coder (Apache 2.0); tldr-pages CC BY 4.0; NL2Bash GPL-3.0.
83
+ Details: https://github.com/arxhr007/lxask/blob/main/NOTICE
84
+
85
+ ## Benchmark
86
+
87
+ InterCode-ALFA results: to be added.