File size: 1,664 Bytes
c019905 44a14a9 c019905 44a14a9 8ce2375 44a14a9 046f9f0 44a14a9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 |
---
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) |