Plain Malay in, shell command out. A LoRA tune of Qwen2.5-Coder-1.5B-Instruct for camne, which runs it locally on CPU — 4 cores, 8 GB, no GPU, nothing leaving the machine.

The point is the register. Machine translation produces "Bagaimanakah cara untuk mencari fail..."; nobody types that. People type camne nak cari file — and more often they type rojak, Malay grammar around English technical nouns, because Malaysians do not translate file, port or server.

InterCode-ALFA, unmodified scorer, 300 tasks per set. Temperature 0, n_predict 64, embedding threshold 0.75, repeat_penalty 1.08, single-line GBNF grammar. Seed 42. Full history, including the runs that failed, in the repo's RESULTS.md.

Paired exact McNemar against that model, 95% CI on the difference:

Read the English row honestly. This revision is worse than the English-only model on advanced English one-liners, by the same 0.06 as the previous revision. Someone who only ever types English is better served by that model. camne exists for the person who types Malay or rojak, and for the first fifty things a beginner asks.

Why this revision shipped. Against the previous revision the benchmark cannot tell them apart (BM +0.010, rojak +0.027, EN 0.000, all inside the 95% CI; the same recipe re-run reproduces to ±0.013). What changed is the shape of the answers on beginner tasks: the previous revision answered create a new file with touch path/to/file1 path/to/file2 ..., create a new folder with skicka mkdir path/to/folder, create a new user with kcadm.sh create users; this one says touch newfile.txt, mkdir newfolder, sudo useradd -m newuser. On the repo's 222-prompt probe: placeholder answers 29 → 2, beginner tasks 0.90 → 0.94 (p = 0.12), held-out tools 0.90 → 0.85 (5 lost, 3 gained). It is a better answer for the person at the prompt, not a higher score.

The pool is 228,357 rows in four registers (formal BM, colloquial, rojak, English), built from NL2SH-ALFA (MIT), tldr-pages (CC-BY-4.0), commandlinefu via b-mc2/cli-commands-explained (CC0) and 0xrushi/git-instruction-dataset (MIT), plus 2,496 hand-written beginner rows (326 tasks: create, list, delete, copy, move a file or folder; read, find, disk, processes, permissions, archives, network, editors, git, packages) with real filenames in all four registers. Rare tools are floored, long-tail tools are capped at 300 rows outside the coreutils, find is held to 5%, and rows whose command carries a path/to placeholder were rewritten to a concrete name on both sides or dropped; nothing is duplicated. Commands are byte-identical to their sources and were never translated — that is what keeps the benchmark scorer usable. The Malay side was machine-translated locally with Gemma-SEA-LION-v3-9B-IT, with Malay technical nouns kept outside the rojak register.

LoRA r32 / α64 / dropout 0.05, all linear layers, 2e-4 cosine with 3% warmup, one epoch, seq 160, effective batch 32, bf16, seed 42, one RTX 3090.

After the SFT stage, one DPO epoch (TRL, β 0.1, lr 5e-6, LoRA r32 on the merged checkpoint, 3,430 pairs, seed 42): chosen = the beginner-task command, rejected = what the SFT model actually said when it was wrong, plus the same command with its names replaced by path/to/... or its tool swapped for an obscure one from the pool. The benchmark tasks are held out of the pairs.

Previous revisions are the commits before this one in this repository's history: run 7 (digest 7576c375…ba67) BM 0.487 / rojak 0.490 / EN 0.543; run 4 (digest 391648d8…4fe8) 0.437 / 0.490 / 0.553, same sets and scorer.

$ camne nak cari file lagi besar dari 100MB kat folder ni
find . -size +100M

Prompt format is ChatML with the system message camne sends:

You are a shell command generator. Output exactly one line: a single POSIX/bash
command that accomplishes the user's request. No prose, no markdown fences, no
explanation.

Decoding must match the benchmark settings above; repeat_penalty 1.08 in particular is not optional. Without it greedy decoding degenerates — nak buat file baru answers : > file.txt instead of touch file.txt.