---
license: mit
pipeline_tag: text2text-generation
tags:
- food
datasets:
- pocasrocas/recipe-gantt
language:
- en
---

## Model Summary

recipe-gantt is a monotask language model to translate recipes into gantt charts in a TSV format.

It is a QLoRA finetune of [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1) on the [pocasrocas/recipe-gantt](https://huggingface.co/datasets/pocasrocas/recipe-gantt) dataset. I then used [llama.cpp](https://github.com/ggerganov/llama.cpp) to convert it to .gguf format for fast local inference.

## Intended use

To create recipe gantt charts and recipe gantt charts only 🧑‍🍳 

It's released here to accompany the [recipe-gantt tool](https://github.com/jbremz/recipe-gantt).

## Input format

The model was trained with the [alpaca](https://github.com/tatsu-lab/stanford_alpaca?tab=readme-ov-file#data-release) instruction/input/response prompt format. Exact details on the expected input can be inferred from the inference code [here](https://github.com/jbremz/recipe-gantt/blob/main/recipe-gantt.py).

## Limitations

- the model will inherit any limitations of the parent model (Mistral-7B-v0.1)
- it was finetuned on a tiny synthetic dataset of only 288 examples so sometimes produces corrupted TSV files and populates cells inaccurately.

## Training

- QLoRA finetune using [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl)
- ~1hr on NVIDIA GeForce RTX 3090 Ti ([wandb](https://wandb.ai/pocasrocas/recipe-gantt/runs/1ostj66y?workspace=user-pocasrocas))
- Training code [here](https://github.com/jbremz/recipe-gantt/tree/720a7e1b3594c03b46209671271fe3c2159fff16/notebooks/002-train)