Modelpublic

C-H-Liu/bash2nl-qwen2.5-coder-3b-GGUF

sourceHugging Faceotherupdated 7d agoView on Hugging Face

1likes69downloads

Model Card

bash2nl — Qwen2.5-Coder-3B QLoRA, GGUF q4KM

Explains a Bash command line in exactly one English sentence, phrased as an instruction starting with a verb. QLoRA fine-tune of Qwen/Qwen2.5-Coder-3B-Instruct, merged and quantized to q4KM.

find . -name "*.py"
-> Find all *.py files/directories under current directory

Usage

bash

hf download C-H-Liu/bash2nl-qwen2.5-coder-3b-GGUF --local-dir bash2nl-3b
cd bash2nl-3b && ollama create bash2nl-3b -f Modelfile
ollama run bash2nl-3b 'ps -ef | grep nginx | awk "{print \$2}" | xargs kill -9'

The Modelfile pins the system prompt and this decoding:

option	value
`temperature`	0.0
`top_k`	1
`top_p`	1.0
`repeat_penalty`	1.0
`num_predict`	96
`num_ctx`	4096
stop	`<	im_end	>`,` <	endoftext	>`

Evaluation

920 held-out commands Judge columns are claude-sonnet-5 scoring a fixed 200-command subset.

metric	3B base	3B qlora f16	3B qlora q4_K_M, no examples	this model	7B qlora q4_K_M
BLEU	13.49	32.04	32.04	32.30	33.49
chrF	43.12	50.85	50.81	50.39	52.21
ROUGE-L	37.58	51.91	51.37	51.19	53.24
judge acceptable+	0.785	0.890	0.825	0.875	0.890
judge wrong	0.215	0.110	0.175	0.125	0.110
avg words	18.5	11.8	11.9	11.9	12.4

License

Apache-2.0, matching the base model, whose weights this GGUF contains. Training data comes from the nl2bash project; consult it for the terms attached to that corpus.