Exploring LLM Priming Strategies for Few-Shot Stance Classification
How does priming affect prompt-based learing? This project aims at analyzing this effect in stance classification. We train a stance classifier on the ibm stance classification dataset by fine-tuning an Alpaca model with a prompt and analyzing how does the selection of the few shots used in the prompt affect the performance of the model.
- To evaluate the prompt-fine-tuning, run the following command
- Hyperparamter optimization
python scripts/run_prompt_fine_tuning.py --validate --optimize
- T evaluate the prompt fine-utning approach using best hyperparameters in the config file.
python scripts/run_prompt_fine_tuning.py --validate
- To evaluate the prompting approaches (prompt) setup run
python scripts/run_prompting.py --validate --optimize
- To optmize DeBERTa using the hyperparameters in the config file, run the following
python scripts/optimize_baseline.py
- run the majority baseline
python baseline.py --vast --majority --offline
- Run prompting or prompt-fine tuning approaches with anaylze k as folows
python scripts/run_prompt_fine_tuning.py --analyze-k
The ks or topic-counts are stored in conf.yaml and the analysis will be run for multiple seeds and the performance will be averaged and saved in the corresponding entry in /bigwork/nhwpajjy. To perform an analysis for one k instnance you can give it as a parameter using --k to produce the results for one seed you can specify the seed using --seed
- Store the path of the produced csv file in conf.yaml under
analyze-k
The results of the experiments will be logged to your home directory. The parameters can be saved in config.yaml
Priming Sampling strategies
To run an experiment with a priming sampling strategy use the parameter --similar-examples. This will retrieve
priming examples tailored toward the test instance. The priming examples are selected using Sentence-Transformers --sentence-transformer.
Example,
python scripts/run_prompting.py --validate --topic-similar --sentence-transformer
To run an experiment with a diversification sampling strategy use the parameter --diverse-examples. This
will takes precomputed cluster centers of the training set as few shots. The number of clusters is then the few shot count
provided in the configuration.
Topic Priming
Topic-priming examples are sampled using a topic similarity, which relies on sentence transformer. the similarities can be used to apply the right sampling strategy.
- To compute the similarities between all the validation and training arguments run the following
python scripts/run_develop_similarity_measure.py --compute-similarity --ibmsc
You have to specify the model used to compute the similarity to be --sentence-transformer
To load similarities from the code you can use the
similarities = load_similarities("ibmsc", "validation", similarity)
which returns a dictionary of dictionary where the indices of the first dictionary are the test indices and the indeices of the nested dictionary for each test index are the indices of the training set with the values being the similarity scores.
To find similar or diverse arguments for an argument in the validation or test set, you can use
examples = sample_diverse_examples(experiment, experiment_type, few_shot_size)
similar can be used for sample_similar
examples = sample_similar_examples(test_istance_index, similarities, df_training, few_shot_size)
Similar examples for alll datasets are cached for efficiency and can be created by running the following commands
python argument_sampling/sampling_strategies.py" --ibmsc
to generate similar examples while using percentiles similarity thresholds you can use
python argument_sampling/sampling_strategies.py" "--ibmsc" --percentiles 10
Stance Priming
Stance-priming approaches use stance similarity to retrieve examples with similar stance to an input instance. To train the contrastive similarity measure, run
python scripts/traing_contrastive_learning.py --vast --lr --epochs --batch-size --output-path-model
The best parameters are also saved in the config file, which will be used as default for each dataset.
To optimize the contrastive similarity measure on the validation set using the hyperparamters in the script, run.
python scripts/traing_contrastive_learning.py --vast --optimize --results-path "$DATA_PATH/contrastive_learning/models/optimization-contrastive-learning-ibmsc.tsv"
To generate the similarity matrix using the contrastive similarity measure use
python scripts/generate_stance_priming_samples.py --vast --similarity-matrix
To generate the stance priming samples using most similar , run the following
python scripts/generate_stance_priming_samples.py --vast --different-topics
to generate similar examples while using percentiles similarity thresholds you can use
python scripts/generate_stance_priming_samples.py" "--ibmsc" --percentiles 10 --different-topics
Diversification strategy
To sample diverse examples, we use ward hierarchical clustering algorithm to cluster the trianing examples into $k$ cluster. The center of each cluster is then taken as an example. To find the diverse examples, we use the following command
python scripts/run_sample_diverse_examples --vast --validate --ibmsc
This will precompute the cluster centers for k in [2, 4, 8, 16, 32, 64]
To sample diverse examples, you can use the following command
examples = sample_diverse_examples(experiment, experiment_type, few_shot_size) where experiment is ibmsc or vast
and experiment_type is validation or test.
Analysis
Mainly there are three types of analyses implemented on the prompting and instruction fine-tuning approaches. For this the results for all ks and the specific model and counts should be in the config file
Few shot size or training topic count effect on priming approaches
python script/run_visualize_over_k_performance.py --k --prompting --prompt-fine-tuning
- Run the signifance tests. For this you need to store the predictions of the models using --path-predictions for all seeds
python notebooks/signifiance_test.py
- Run the prime anaylis for stance priming you can run
python "$CODE_PATH/scripts/run_prompting.py" "--alpaca-7b" "--ibmsc" --offline --vllm --analyze-prime-similarity \
--similar-examples --path-similar-examples "$DATA_PATH/sampling-strategies/ibmsc-similar-stance-examples.tsv"
- Run the prime anaylis for topic priming you can run
python "$CODE_PATH/scripts/run_prompting.py" "--alpaca-7b" "--ibmsc" --offline --vllm --analyze-topic-similarity \
--similar-examples --path-similar-examples "$DATA_PATH/sampling-strategies/ibmsc-similar-stance-examples.tsv"
1 maintainer
Metadata
1 maintainer
Release files for few-shot-priming 0.3.950
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| few_shot_priming-0.3.950.tar.gz | 202.5 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| few_shot_priming-0.3.950-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 410.2 kB
Release files / few_shot_priming-0.3.950.tar.gz
| Download URL | few_shot_priming-0.3.950.tar.gz |
|---|---|
| Size | 202.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
56d303e9eb100557ddcb6704cb2273f94ac36b43b2bf6e55ec192c480bba284b
|
|
BLAKE2b-256 checksum How to use checksums |
6e267d2be3a175cb42ccd09c3a0a1360d99e92036f610ce3e37e949f2f2e2304
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.11.8
|
Release files / few_shot_priming-0.3.950-py3-none-any.whl
| Download URL | few_shot_priming-0.3.950-py3-none-any.whl |
|---|---|
| Size | 207.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
38124d488c709eba821b20caea17cdf035aea3d221d4316151f32cff81f08209
|
|
BLAKE2b-256 checksum How to use checksums |
b7530124fac8d72793127b5dddf1c02c267b6c19fe97087f0b1a023d049f119c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.11.8
|
