Data Curation Alone Can Stabilize In-context Learning (ACL 2023)
Ting-Yun Chang and Robin Jia
Abstract: In-context learning (ICL) enables large language models (LLMs) to perform new tasks by prompting them with a sequence of training examples. However, it is known that ICL is very sensitive to the choice of training examples: randomly sampling examples from a training set leads to high variance in performance. In this paper, we show that carefully curating a subset of training data greatly stabilizes ICL performance without any other changes to the ICL algorithm (e.g., prompt retrieval or calibration). We introduce two methods to choose training subsets -- both score training examples individually, then select the highest-scoring ones. CondAcc scores a training example by its average dev-set ICL accuracy when combined with random training examples, while Datamodels learns linear regressors that estimate how the presence of each training example influences LLM outputs. Across five tasks and two LLMs, sampling from stable subsets selected by CondAcc and Datamodels improves average accuracy over sampling from the entire training set by 7.7% and 6.3%, respectively. Surprisingly, the stable subset examples are not especially diverse in content or low in perplexity, in contrast with other work suggesting that diversity and perplexity are important when prompting LLMs.
This repository is based on MetaICL
- Quick Start
- CondAcc
- Datamodels
- Baselines
- Evaluation
- Data
- Construct
$\mathcal{D}_{\text{ICL}}$ - How to use the released
$\mathcal{D}_{\text{ICL}}$ ? - Stable Subset Examples
$ pip install -r requirements.txt
-
$ bash demo/download_dicl.sh
will download the released prompt-output pairs- see the Construct
$\mathcal{D}_{\text{ICL}}$ section below for more details
- see the Construct
- The proposed CondAcc method is implemented in
select_condacc.py
- To reproduce the results in the paper, see Evaluation
- To train datamodels, run:
$ bash scripts/train_datamodels.sh
- To test datamodels (Appendix A3), run:
$ bash scripts/test_datamodels.sh
- The Datamodels selection is implemented in
select_datamodels.py
- To reproduce the results in the paper, see Evaluation
- Download pretrained datamodels
out_datamodel.zip
- The Oneot baseline:
- First, run 1-shot ICL by
$ bash scripts/run_oneshot.sh {gpu_i}
- The subset selection process is implemented in
baseline_oneshot.py
- First, run 1-shot ICL by
- The TopPrompts baseline is implemented in
baseline_top_prompts.py
- The Calib baseline is implemented in
calib_evaluate.py
$ bash scripts/run_calibration.sh {gpu_i}
- To reproduce the results in the paper, see Evaluation
The following scripts will run the two proposed methods and all baselines.
$ bash scripts/run_label.sh {gpu_i}
$ bash scripts/run_unlabel.sh {gpu_i}
$ bash scripts/run_ood.sh {gpu_i}
- We include the data in
data/
. The files are organized as follow:
data
โโโ glue-sst2
โ โโโ *train.jsonl
โ โโโ *dev.jsonl
โ โโโ *test.jsonl
โ โโโ unlabeled
โ โโโ *train.jsonl
โ โโโ is_groundtruth.npy
โโโ boolq/
โโโ subj/
โโโ scicite/
โโโ glue-mnli/
โโโ ag_news/
- Each task folder contains
*train.jsonl, *dev.jsonl, *test.jsonl
, the gold-labeled train/dev/test splits, and the unlabeled training setunlabeled/*train.jsonl
. - Note that both the labeled and unlabeled setups use the same dev sets for method developement and are evaluated on the same test sets.
- To reproduce the data creation for the unlabeled setups, see
create_unlabeled.py
- We release the set of prompt-output pairs
$\mathcal{D}_{\text{ICL}}$ inDicl
. The files are organized as follow:
Dicl
โโโ gpt-j-6b
โ โโโ label_glue-sst2
โ โ โโโ *train_ids.npy
โ โ โโโ *permute_ids.npy
โ โ โโโ *sampled.pkl
โ โ โโโ merged*.pt
โ โโโ unlabel_glue-sst2/
โ โโโ label_boolq/
โ โโโ unlabel_boolq/
โ โโโ label_subj/
โ โโโ unlabel_subj/
โ โโโ label_scicite/
โ โโโ unlabel_scicite/
โ โโโ label_ag_news/
โ โโโ unlabel_ag_news/
โโโ opt-13b/
โโโ opt-6.7b/
โโโ gpt-neo-2.7B/
-
Each task has three folders, e.g.
label_glue-sst2
andunlabel_glue-sst2
for labeled/unlabeled setup (Sec 2), andtest_glue-sst2
for evaluating how well Datamodels can approximate the target LLM on the held-out prompts (Appendix A3). -
Each task folder contains 4 files:
-
*sampled.pkl
: the list of sampled prompts, where each prompt consists of a list of$K$ training examples. -
*train_ids.npy
: the training example IDs in each prompt, where each row in the array consists of$K$ example IDs. -
*permute_ids.npy
: the permutation IDs of each prompt. -
merged*.pt
: given the prompts, the LLM's output logits before softmax.
-
-
To reproduce the construction of
$\mathcal{D}_{\text{ICL}}$ :-
$ bash scripts/build_dicl.sh {gpu_i} {segment_i} {permute_i}
-
gpu_i
: the available gpu id, default 0. -
segment_i
: [0-4]. We divide the list of sampled prompts into 5 segments to run them in parallel. -
permute_i
: [0-1]. Given the same prompt, we run 2 different permutations. - add the argument
--is_unlabel
to build$\mathcal{D}_{\text{ICL}}$ for the unlabeled setup.
-
- Note: the construction process may take hundreds of GPU hours for a task
-
- We include an example code in
demo
-
$ bash demo/download_dicl.sh
will automatically download and unzipDicl.zip
-
$ python -m demo.demo_dicl --model gpt-j-6b --task glue-sst2
, use--is_unlabel
for the unlabeled setup.- this will return a list of datapoints, where a datapoint is a dict that looks like:
-
{'train_examples': [{ 'input': 'Review: whole mess \nSentiment:', 'options': ['negative', 'positive'], 'output': 'negative', 'task': 'glue-sst2'}, { 'input': 'Review: but it also comes with the ' 'laziness and arrogance of a thing that ' "already knows it 's won . \n" 'Sentiment:', 'options': ['negative', 'positive'], 'output': 'negative', 'task': 'glue-sst2'}, { 'input': 'Review: intelligent and moving . \n' 'Sentiment:', 'options': ['negative', 'positive'], 'output': 'positive', 'task': 'glue-sst2'}, { 'input': 'Review: does point the way for ' 'adventurous indian filmmakers toward a ' 'crossover into nonethnic markets . \n' 'Sentiment:', 'options': ['negative', 'positive'], 'output': 'positive', 'task': 'glue-sst2'}], 'train_ids': np.array([101, 286, 666, 623]), 'dev accuracy': 0.85, 'logits': a torch.FloatTensor of shape [n_dev, n_labels]}
- We include the identified stable subset examples in
out_select
label_stable_subsets/{model}-{task}-{CondAcc/Datamodels}.jsonl
each file shows a stable subset (20 examples) identified by CondAcc/Datamodels in the labeled setupunlabel_stable_subsets/{model}-{task}-CondAcc.jsonl
each file shows a stable subset (20 examples) identified by CondAcc in the unlabeled setupgood_example_ids/*.npy
each file shows the corresponding 20 example IDs