Userb1az/Mistral-7B-Instruct-v0.3-DockerNL-sft

Mistral-7B-Instruct-v0.3 Docker NL (LoRA)

Overview

This repository contains a Mistral-7B-Instruct-v0.3 model fine‑tuned to translate natural language requests about Docker into single Docker CLI commands.
Fine‑tuning was performed via LoRA using MattCoddity/dockerNLcommands.

Typical behavior:

• Input: Give me the Docker images that were built before the image "nginx:latest".
  Output: docker images -f before=nginx:latest

• Input: Give me a list of containers that have the Ubuntu image as their ancestor.
  Output: docker ps -a --filter 'ancestor=ubuntu'

• Input: Please show me a table of images, including the repository, tag, ID, and size.
  Output: docker images --format "{{.Repository}},{{.Tag}},{{.ID}},{{.Size}}"

The LoRA adapter was merged into the base Mistral weights using LLaMA‑Factory, so this repo exposes a standard Hugging Face Transformers checkpoint.

Base model

• Base: mistralai/Mistral-7B-Instruct-v0.3
• Architecture: decoder‑only, 7B parameters
• Intended use: instruction‑following in multiple languages

Training details

• Framework: LLaMA‑Factory (SFT / LoRA)
• Finetuning type: LoRA, merged into full model
• Dataset: MattCoddity/dockerNLcommands (train split)
• Task: single‑turn instruction tuning (NL Docker request → one Docker command)
• Max sequence length: 2048 tokens
• Epochs: 3
• Batch / grad accumulation: effective batch size 64
• Learning rate: 5e‑5 (AdamW, cosine scheduler)
• Precision: bfloat16

Loss at the end of training was around 0.10, indicating good fit on the narrow task.

For best results:

• Use temperature=0.0–0.1 and max_new_tokens around 32–64.
• Treat the model as a command generator: always review commands before running them in production shells.

Limitations

• The model is trained on a relatively small, Docker‑focused dataset and should not be treated as a general‑purpose assistant.
• It may produce incorrect or unsafe commands in edge cases; human review is required.