AI Tools.

Search

text generation by facebook

opt-1.3b

OPT-1.3B is Meta's Open Pre-trained Transformer at 1.3 billion parameters, released in 2022 as part of a suite ranging from 125M to 175B. The model was trained on a curated mix of publicly available datasets and released with full weights and training logs to enable reproducibility research. It has largely been superseded by later open LLMs but remains a useful controlled baseline.

Summary text generated by an automated pipeline from the model card · Not individually reviewed or run by us · How this page is made

From the model card

Fields below are copied from the tags and counters on the HuggingFace repository facebook/opt-1.3b at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.

Publisher (HF namespace)
facebook
Pipeline tag
text-generation
Library
Transformers
Framework tags
PyTorch, TensorFlow, JAX
License tag
other — read the license file in the repo before relying on it
Language tags
English (en)
Papers cited
arXiv:2205.01068, arXiv:2005.14165
Downloads (HF counter at last fetch)
352,352
Likes (HF counter at last fetch)
184
Model card
https://huggingface.co/facebook/opt-1.3b

Use cases

  • Reproducing OPT-paper ablations on architecture and data choices
  • Low-memory text generation for researchers without large GPU clusters
  • Comparative studies on pre-2023 LLM training practices
  • Teaching language model concepts with a well-documented checkpoint
  • Baseline in continual learning and catastrophic forgetting experiments

Pros

  • Full training logs and logbook released alongside weights for transparency
  • PyTorch, TF, and JAX weights available
  • 1.3B parameters run feasibly on 4GB VRAM or CPU
  • Extensively benchmarked in the literature for easy comparison

Cons

  • Significantly outperformed by Llama 2 7B and its descendants on every benchmark
  • 'Other' license with Meta's terms limits some commercial deployments
  • No instruction tuning; continuation model with poor instruction following
  • Training data biases are documented in the OPT paper and carry real risks

Tags

transformerspytorchtfjaxtext-generationoptenarxiv:2205.01068arxiv:2005.14165license:otherdeploy:azureregion:us