ltgoslo / norbench Goto Github PK

There is a set of North-T5 models that are specifically trained on modern Norwegian Bokmål/Nynorsk. Run all the benchmarks on the following models:

• https://huggingface.co/north/t5_small_NCC_modern
• https://huggingface.co/north/t5_base_NCC_modern
• https://huggingface.co/north/t5_large_NCC_modern

subj

https://huggingface.co/docs/datasets/create_dataset

After merging #18, we again face CUDA out-of-memory errors. The data is loaded successfully, the LM is also loaded, the training attempts to start, and then:
RuntimeError: CUDA out of memory. Tried to allocate 12.00 MiB (GPU 0; 23.70 GiB total capacity; 1.10 GiB already allocated; 10.56 MiB free; 1.18 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

This happens even with extremely small batch sizes and even on GPUs with 24GB and more RAM. This is the command I am running:
python3 norbench_run.py --path_to_model XXX --task sentiment --task_specific_info sentence --model_name XXX --batch_size 2 --epochs 5

I tried again with the (now deleted) sentiment_analysis/finetuning.py script:
python3 finetuning.py -level sentence -model XXX -batch_size 16 -epochs 5
It works like a charm on the same data, same hardware and same set of Python modules.

Thus, obviously something is wrong with some calls in the norbench_run.py script. Please fix it.

Hi!
Was just wondering if there's any open-source license that applies to the code in the repo
or if it was a conscious decision to leave one out.
I understand that different licenses apply to the different datasets in the repo. Perhaps it could be listed in the readme?
It would be nice to know what we can and can't do with the code

The current issue provides the information about implementing XLM-R for the tasks mentioned below:

1. Part-of-Speech tagging Bokmål
2. Part-of-Speech tagging Nynorsk
3. Binary sentiment analysis
4. Named entity recognition Bokmål
5. Named entity recognition Nynorsk

This issue provides information about the baseline and models that have been tested on the data of the Wordnet from the Norwegian National Library.

Models that have already been tested for provided data:

1. Most frequent sense baseline (with the using of Norbert2)

2. A number of tests with different input data on Norbert2, Multilingual Bert (12 models are evaluated by the moment)

Other tests are in the process of implementation.

Planned updates for TSA evaluation script:

• Put trust_remote_code in model_args
• Pull dataset from Huggingface in stead of local copy
• Put all needed helper functions and classes into the tsa_finetuning script

After this, tsa_finetuning.py should not depend on any other files in the repo.

except ner_eval which also is used by the NER script

Analogous to https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard etc.

Hi!

I'm currently researching open LLMs for Norwegian, and since a lot of SOTA solutions nowadays are based on scaling up
GPT-style causal LM models and fine-tune them on instructions, I think it would make sense to have
some evaluation and fine-tuning scripts and metrics for these types of models too.

I've already made a fork where I'm working on this myself, but I'm wondering:

1. Is this something that is already being looked into?
2. If not, could a contribution to this main repo be desirable?

Running evaluation on all tasks crashes on the very first one with:

Traceback (most recent call last):
  File "norbench/evaluation_scripts/all_tasks.py", line 134, in <module>
    run_models_for_current_task(do_train, tsk, name_sub_info, path_tsk, run_name, model_identifier, epochs, use_seqeval_evaluation_for_ner, use_class_weights_for_sent)
  File "norbench/evaluation_scripts/all_tasks.py", line 82, in run_models_for_current_task
    run_tasks(do_train, current_task, name_sub_info, data_path, model_identifier, run_name, epochs, use_seqeval_evaluation_for_ner, use_class_weights_for_sent)
  File "norbench/evaluation_scripts/all_tasks.py", line 55, in run_tasks
    dev_score =  tasks[current_task]['eval'](data_path, "dev", sub_task_info=name_sub_info, short_model_name=model_identifier, task=current_task)
  File "norbench/evaluation_scripts/pos_finetuning.py", line 37, in test
    test_lang_path = data_path + task
TypeError: unsupported operand type(s) for +: 'bool' and 'str'

The command is
python3 all_tasks.py --path_to_model MODEL --task all --download_cur_data True --model_name IDENTIFIER

The current issue provides information about the models that have been implemented within the framework of Norbench.

The information provided below will contain:

• documentation for running scripts for the current (POS-tagging, Binary Sentiment Analysis, NER) task
• list of models that can be used for each of the tasks (at the current moment)
• future updates, new available models, etc.

The download_cur_data argument in the all_tasks.py script behaves a bit weirdly.

1. A user reasonably expects it's enough to simply append --download_cur_data to the command, and the datasets will be downloaded automatically. Instead, the script crashes with
   all_tasks.py: error: argument --download_cur_data: expected one argument

2. If instead --download_cur_data True is used, the script still crashes, this time with

File "all_tasks.py", line 120, in <module>
    raise Exception(f'Check paths "{path_tsk}" to data for {tsk} task')
Exception: Check paths "" to data for pos task

I believe boolean arguments should be implemented with BooleanOptionalAction.

This issue contains information about statistics in the Norwegian Wordnet (Bokmål) from the National Norwegian Library.

At the initial stage, statistics was made on the official dataset from the National Library in general, showing distribution of number of examples, pos tags, and senses per lemma.

At this stage, more detailed statistics are carried out, including the following points:

1. the statistics of distribution of unique sentences for the given lemma: choice was concentrated on lemmas that could provided 5 or more sentences through the dataset.

2. it is proposed to divide words into categories depending on the number of possible senses.