forked from KEMT/zpwiki
132 lines
4.4 KiB
Markdown
132 lines
4.4 KiB
Markdown
---
|
|
title: Youssef Ressaissi
|
|
published: true
|
|
taxonomy:
|
|
category: [iaeste]
|
|
tag: [summarization,nlp]
|
|
author: Daniel Hladek
|
|
---
|
|
|
|
|
|
IAESTE Intern Summer 2025, 1.7. - 31.8.2025
|
|
|
|
Goal: Evaluate and improve language models for summarization in Slovak medical or legal domain.
|
|
|
|
|
|
Tasks:
|
|
|
|
1. Get familiar with basic tools
|
|
- and prepare working environment: HF transformers, datasets, lm-evaluation-harness, HF trl
|
|
- Read several recent papers about summarization using LLM and write a report.
|
|
- Get familiar how to perform and evaluate document summarization using language models in Slovak.
|
|
2. Make a comparison experiment
|
|
- Pick summarization datasets and models. Evaluate several models for evaluation using ROUGE and BLEU metrics.
|
|
- https://github.com/slovak-nlp/resources
|
|
- Describe the experiments. Summarize results in a table. Describe the results.
|
|
3. Improve performance of a languge model.
|
|
- Use more data. Prepare a domain-oriented dataset and finetune a model. Maybe generate artificial data to imporve summarization.
|
|
- Run new expriments and write down the results.
|
|
4. Report and disseminate
|
|
- Prepare a final report with analysis, experiments and conclusions.
|
|
- Publish the fine-tuned models in HF HUB. Publish the paper from the project.
|
|
|
|
Meeting 19.8.
|
|
|
|
State:
|
|
|
|
- Fine tuned Slovak Mistral 7B
|
|
- Tried Llama3 7B - results look ok, but MIstral is Better.
|
|
- Tried gpt-oss, but it does not work because of dependencies.
|
|
- Work on preliminary final report.
|
|
- ROUGE score is not good for abstractive summarization.
|
|
- The best way to evaluate so far is to see it in person.
|
|
|
|
Tasks:
|
|
|
|
- Try to fine tune other models. 'google/gemma-3-4b-it, HPLT/hplt2c_slk_checkpoints Qwen/Qwen3-4B'. Results wil be in different branches of the repository.
|
|
- Try to automatically evaluate the results using a large LLM. Read some papers about it. Prepare a script using ollama and gpt-oss-20B.
|
|
- Work on the final report.
|
|
|
|
|
|
|
|
Meeting 4.8.
|
|
|
|
State:
|
|
|
|
- Tested LMs with ROUGE metrics, most models got 4-5 ROGUE, facebook/mbart-large-50 got 17 (trained for translation).
|
|
- In my opinion, large-50 is not good for finetuning, because it is already fine tuned for translation.
|
|
- no finetuning done yet.
|
|
|
|
Tasks:
|
|
|
|
- Try evaluate google/flan-t5-large, kiviki/mbart-slovaksum-large-sum and similar models. These should be already working.
|
|
- continue working on finetuning t5 or Mbart models, but ask when you are stuck. Use hf examples script on summarization
|
|
|
|
Future tasks:
|
|
|
|
- use LLMS (open or closed) and evaluate (ROUGE) summarization without fine-tuning on slovak legal data set
|
|
- install lm-eval-harness, learn it, prepare and run task for slovak summarization
|
|
|
|
|
|
Meeting 13.8.2025
|
|
|
|
State:
|
|
|
|
- Managed to fine-tune Slovak Mistral-7B on Legal documents. One epoch is enough.
|
|
- Ditched T5 pipeline
|
|
|
|
Tasks:
|
|
|
|
- try to fine tune some different llm and compare the results.
|
|
- try to fine tune with the news dataset.
|
|
- prepare a table - tables with results
|
|
- compare with zero- shot scenario (with various models)
|
|
- dont forget to put scripts on GIT.
|
|
|
|
Future task:
|
|
|
|
- Find the optimal hyperparameters.
|
|
- Write the technical report, where you summarize methods, experiments and results.
|
|
|
|
|
|
Meeting 24.7.
|
|
|
|
State:
|
|
|
|
- Working custom environment with JupyterNotebook
|
|
- Fine-tuning mbart, results are not great
|
|
|
|
Tasks:
|
|
|
|
- Try T5 based models: slovak t5-base, umt5, mt5, flan-t5,
|
|
- Zero shot-evaluate LLMs on news and legal data.
|
|
- Find a way to fine-tune LLM for summarization.
|
|
- Fine-tune LLM for summarization.
|
|
|
|
Meeting 17.7.2025:
|
|
|
|
State:
|
|
|
|
- Studying of the task, metrics (ROUGE,BLEU)
|
|
- Loaded a model. preprocessed a dataset, evaluated a model
|
|
- loaded more models, used SlovakSum, generated summarization with four model and compare them with ROUGE and BLEU (TUKE-KEMT/slovak-t5-base, google/mt5-small, google/mt5-base, facebook/mbart-large-50)
|
|
- the comparison is without fine tuning (zero shot), for far, the best is MBART-large
|
|
- working on legal dataset "dennlinger/eur-lex-sum",
|
|
- notebooks are on the kemt git
|
|
|
|
Tasks:
|
|
|
|
- Prepare "mango.kemt.fei.tuke.sk" workflow
|
|
- Finetune an existing models and evaluate it. Use News and Legal datasets
|
|
- Try mbart-large, flan-t5-large, slovak-t5-base, google/t5-v1_1-large
|
|
- Describe the experimental setup, prepare tables with results.
|
|
|
|
|
|
Future tasks:
|
|
|
|
- Try prompting LLM and evaluation of the results. We need to pick LLM with SLovak Support
|
|
- Finetune an LLM to summarize
|
|
- Use medical data (after they are ready).
|
|
- Prepare a detailed report (to be converted into a paper).
|
|
|