A two-model system for reasonable text generation (via vector scoring)
"Think before you speak"
This repo demonstrates how to use two language models (LMs) to achieve more lucid and coherent text generations.
...these best scoring texts are given output priority:
...these worst scoring ones get filtered out:
Causal-LMfor text generation (e.g.distilgpt2)Masked-LMfor generated text criticism (e.g.distilroberta)
- Select a specific domain to build generators for (e.g. Machine Learning Ideas)
- Acquire a text corpus for the domain at hand (e.g.
aalksii/ml-arxiv-papers) - Fine-tune the
Causal-LMon the corpus (e.g. finetune_causal.ipynb) - Fine-tune the
Masked-LMon the corpus (e.g. finetune_masked.ipynb) - Acquire and check quality of
Masked-LMvectors on the corpus and save them (e.g. vectors.ipynb) - Determine the generation objective of the
Causal-LMw.r.t. the embeddings of theMasked-LM; for example:- Novelty: generated ML idea should be at least
0.05cosine distance away from any existing idea vector - Feasibility: generated ML idea should not be too isolated; it should have at least
10neighbors within0.1cosine distance away
- Novelty: generated ML idea should be at least
- Generate texts from
Causal-LM; only output those that pass the objective. (e.g. generation.ipynb)
There is potential for Reinforcement Learning-inspired improvements that could be made here.