TL;DR: A deterministic Phi-3 fine-tune that converts natural language into a single, canonical, POSIX-safe Bash command — no explanations, no pipelines.

phi3-nl2bash-canonical

Model Summary

phi3-nl2bash-canonical is a task-specialized small language model fine-tuned to translate natural-language instructions into minimal, valid Bash commands.
The model is intentionally constrained to produce single, canonical, POSIX-safe commands without explanations, pipelines, subshells, or side effects.

This model is designed for command-line education, tooling, and evaluation, not for general-purpose chat.

Base Model

microsoft/phi-3-mini-4k-instruct

Training Data

The model was fine-tuned using a curated subset of the NL2Bash dataset combined with synthetic examples generated from a manually verified command core.

Dataset characteristics:

Only local, single-command Bash instructions
No pipelines, redirections, subshells, SSH, rsync, or environment variables
Restricted command set (e.g., ls, cd, mkdir, touch, cp, mv, chmod, cat, head, tail, basename, dirname, wc)
Synthetic augmentation used to improve coverage while preserving canonical form

The goal was precision and determinism, not breadth.

Training Method

Parameter-efficient fine-tuning (LoRA)
Conservative hyperparameters to avoid catastrophic forgetting
Instruction format: ChatML-style (<|user|>, <|assistant|>, <|end|>)

Model Variants

This repository contains multiple formats:

FP16 GGUF (gguf/fp16/):
Canonical archival format for reproducibility and re-quantization
Q4_0 GGUF (gguf/q4_0/):
CPU-efficient quantized model suitable for local inference (e.g., Ollama, llama.cpp)

Usage

Ollama

Example Modelfile:

FROM ./phi3-nl2bash-canonical-17012026.q4_0.gguf

SYSTEM You output only valid bash commands.
SYSTEM No explanations or markdown.

TEMPLATE """<|user|>
{{ .Prompt }}
<|assistant|>
"""

PARAMETER stop "<|end|>"
PARAMETER temperature 0

Example

Prompt: create a file called a

Output: touch a

Intended Use

Teaching command-line basics
Evaluating NL→CLI translation
Safe, constrained automation

Out of Scope

Complex shell scripting
Remote execution
File discovery or destructive commands
General-purpose conversation

Limitations

This model intentionally trades expressiveness for safety and determinism.
It may refuse or oversimplify complex requests.

Ethics & Safety

The model was trained to avoid unsafe shell constructs and does not generate commands involving networking, privilege escalation, or destructive operations unless explicitly specified in the prompt.

Citation

If you use this model in academic work, please cite the accompanying repository.Developed as part of an academic thesis (2026).

Downloads last month: 160

Safetensors

Model size

4B params

Tensor type

BF16