|
--- |
|
license: openrail |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
**To use MathT5 easily:** |
|
|
|
1. Download ```MathT5.py```. |
|
2. ```from MathT5 import load_model, inference``` |
|
3. ```tokenizer, model = load_model("jmeadows17/MathT5-large")``` |
|
4. ```inference(prompt, tokenizer, model)``` |
|
|
|
```MathT5.pretty_print(text, prompt=True)``` makes prompts and outputs (```prompt=False```) easier to read. |
|
|
|
**Overview** |
|
|
|
MathT5-large is a version of FLAN-T5-large fine-tuned for 25 epochs on 15K (LaTeX) synthetic mathematical derivations (containing 4 - 10 equations), that were generated using a symbolic solver (SymPy). |
|
It outperforms the few-shot performance of GPT-4 and ChatGPT on a derivation generation task in ROUGE, BLEU, BLEURT, and GLEU scores, and shows some generalisation capabilities. |
|
It was trained on 155 physics symbols, but struggles with out-of-vocabulary symbols. Paper available here: https://arxiv.org/abs/2307.09998. |
|
|
|
|
|
**Example prompt:** |
|
|
|
```prompt = "Given \\cos{(q)} = \\theta{(q)}, then derive - \\sin{(q)} = \\frac{d}{d q} \\theta{(q)}, then obtain (- \\sin{(q)})^{q} (\\frac{d}{d q} \\cos{(q)})^{q} = (- \\sin{(q)})^{2 q}"``` |
|
|
|
Output derivations are equations separated by "and". |
|
|
|
Additional prompts can be found in "training_prompts.json" alongside the model files. |
|
|
|
Use ```"jmeadows17/MathT5-base"``` for the lightweight version. |
