OPT-IML (OPT + Instruction Meta-Learning) is a set of instruction-tuned versions of OPT, on a collection of ~2000 NLP tasks gathered from 8 NLP benchmarks, called OPT-IML Bench.
You can use this model directly with a pipeline for text generation.
1>>> from transformers import pipeline
2
3>>> generator = pipeline('text-generation', model="facebook/opt-iml-max-1.3b")
4
5>>> generator("What is the capital of USA?")
While OPT-IML models outperform baseline OPT on an extensive set of evaluations,
nevertheless, they are susceptible to the various risks associated with using large language models
relating to factual correctness, generation of toxic language and enforcing stereotypes. While we release our
OPT-IML models to proliferate future work on instruction-tuning and to improve the availability
of large instruction-tuned causal LMs, the use of these models should be
accompanied with responsible best practices.
OPT-IML models are trained on OPT-IML Bench, a large benchmark for Instruction MetaLearning (IML) of 2000 NLP tasks consolidated into task categories from 8 existing benchmarks include Super-NaturalInstructions, FLAN, PromptSource, etc.
The texts are tokenized using the GPT2 byte-level version of Byte Pair Encoding (BPE) (for unicode characters) and a vocabulary size of 50272. The inputs are sequences of 2048 consecutive tokens.
The 30B model was fine-tuned on 64 40GB A100 GPUs. During fine-tuning, models saw approximately 2 billion tokens, which is only 0.6% of the pre-training
budget of OPT.
1@misc{iyer2022opt,
2 title={OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalization},
3 author={Iyer, Srinivasan and Lin, Xi Victoria and Pasunuru, Ramakanth and Mihaylov, Todor and Simig, D{\'a}niel and Yu, Ping and Shuster, Kurt and Wang, Tianlu and Liu, Qing and Koura, Punit Singh and others},
4 year={2022},
5 eprint={2212.12017},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL}
8}
