llm experiments - Githubissues

akon1te / dynamic-topic-modeling

Final qualifying work of the HSE NN on the topic Dynamic topic modeling for text and audio dialogues

0 stars 0 forks source link

llm experiments #1

Open akon1te opened 5 months ago

akon1te commented 5 months ago

Dataset preparing step

[x] Extract topics from tigae
[x] Extract topics from superseg
[ ] Extract topics from dailydialog

Audio datasets preparing for test stage

[ ] Prepare QMSum corpus

Text extraction from audio files step

[ ] Wave2vec
[ ] Whisper

Clustering step

[ ] Adapt the best ready-made solutions from github to our solution Improving Unsupervised Dialogue Topic Segmentation with Utterance-Pair Coherence Scoring
[ ] Experiment with GitHub - zhang-yu-wei/ClusterLLM: LLM guided text clustering
[ ] Research adaptation of approaches from usual text clustering

Topic extraction step

[ ] Fine-tune LLM on non-dialog data for topic extraction
[ ] Fine-tune LLM on dialogs for topic extraction (Tiage, Superseg)
[ ] Experiment with mixture of dialog and text datasets for training
[ ] Research other approaches

Topics list evolution step

[ ] Research llm approaches
[ ] Research other approaches

akon1te commented 3 months ago

[x] Prepare scripts for prompt generating

0shot

[x] 0shot phi-2
[x] 0shot StableLM
[x] 0shot Gemma

FT

[ ] Fine-tuning phi-2
[ ] Fine-tuning StableLM
[ ] Fine-tuning Gemma

Prepare scripts for LLM inference

[ ] Prepare scripts for LLM inference