Edit model card
AI2 ImpACT License – Low Risk Artifacts (“LR Agreement”) and completing all fields below. This model is licensed under the LR Agreement.

\n","classNames":"hf-sanitized hf-sanitized-oEV558BAb3ewnfJkv3QjC"},"gated":"auto","isLoggedIn":false,"repoId":"allenai/paloma-1b-baseline-pile","repoType":"model","isGoogleGemma":false}">

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access to this model is automatically granted upon accepting the AI2 ImpACT License – Low Risk Artifacts (“LR Agreement”) and completing all fields below. This model is licensed under the LR Agreement.

Log in or Sign Up to review the conditions and access this model content.

Configuration Parsing Warning: In UNKNOWN_FILENAME: "auto_map.AutoTokenizer" must be a string

Model Card for Model ID

The Paloma 1B baselines are a collection of language models pretrained on popular corpora while controlling all other experimental variables. These models are developed to facilitate scientific comparisons of language model fit using the Paloma benchmark of 585 textual domains. This collection of models includes 6 baseline 1B parameter models each trained on ~150B tokens from one the following corpora: Dolma, The Pile, RedPajama, Falcon-RefinedWeb, C4, and MC4-en.

Model Details

Model Description

  • Developed by: Ian Magnusson, Akshita Bhagia, Valentin Hofmann, Luca Soldaini, Ananya Harsh Jha, Oyvind Tafjord, Dustin Schwenk, Evan Pete Walsh, Yanai Elazar, Kyle Lo, Dirk Groeneveld, Iz Beltagy, Hannaneh Hajishirzi, Noah A. Smith, Kyle Richardson, and Jesse Dodge
  • Model type: Decoder-only transformer language model
  • Language(s) (NLP): English
  • License: AI2 ImpACT License – Low Risk Artifacts (“LR Agreement”)

Uses

Direct Use

This model is primarily intended as research artifact that is a baseline for the language modeling benchmark Paloma.

Out-of-Scope Use

The restrictions to use of this model are described in the model license: AI2 ImpACT License – Low Risk Artifacts (“LR Agreement”)

Bias, Risks, and Limitations

This model is purely trained as an autoregressive language model. It has not been adapted in any way to prevent bias. It is a model of the language distribution that it is trained on.

Recommendations

This research artifact is a baseline for a language modeling benchmark. Best uses of this model will take advantage of the experimental controls applied to this model and the other Paloma baselines. These enable comparisons of models that vary only in the pretraining corpus used to train them.

How to Get Started with the Model

Install the code needed to run inference with the model

pip install ai2-olmo

Download and instantiate the model

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("allenai/<model name here>", trust_remote_code=True)

Training Details

Training Data

Each of the Paloma baseline models are trained on one of the following datasets: Dolma, The Pile, RedPajama, Falcon-RefinedWeb, C4, and MC4-en.

Training Procedure

Preprocessing

We remove any document in the pretraining data that is contaminated with respect to the Paloma evaluation data. We match overlaps of evaluation and train text at the paragraph level, i.e., newline separated spans of text. To avoid coincidental collisions in the space of small strings, we ignore matches in paragraphs smaller than 13 unicode segmented tokens. Similarly, we ignore paragraphs composed of only punctuation, spaces, and emoji, as, unlike words, these can be arbitrarily repeated when used as formatting, leading to high frequency n-grams greater than our 13-gram threshold. Lastly, as code data consists almost entirely of short and often repeated lines, we forgo any decontamination against the code evaluations in Paloma.

Training Hyperparameters

The Paloma baseline 1B parameter models that we train employ the following architecture: 2048 maximum sequence length, 2048 model dimension, 16 layers, 16 attention heads, RoPE embedding, SwiGLU activation, mixed precision, non-parametric layer normalization, and sequential model blocks for attention and feed-forward networks. We use EleutherAI's GPT NeoX tokenizer but add 3 additional special tokens that are used to mask PII in Dolma. We train to 35k steps (∼150B tokens) with the following LionW optimizer configurations: 2.0e-4 peak learning rate, warm-up of 2000 steps, cosine decay to 70k steps (∼300B tokens), 0.1 weight decay, and betas of 0.9 and 0.95. Note that our batch size varies slightly to accommodate two groups of baselines that were run on different hardware. The Dolma and Falcon-RefinedWeb baselines were run with a batch size of 2112 training instances per step on 24 A100s. The RedPajama, The Pile, C4, and mC4-EN baselines were run with a batch size of 2048 on 64 AMD Instinct MI250X GPUs. In each case we save model checkpoints every 5k steps (∼20B tokens).

Evaluation

Testing Data, Factors & Metrics

Testing Data

The Paloma benchmark is used to evaluate these baseline models.

Factors

Paloma evaluates on 585 domains. These are a collection of the most fine-grained domains readily available in current metadata.

Metrics

Paloma measures langauge modeling fit using Perplexity. It is a benchmark of language modeling, so examination of downstream uses is out of scope.

Results

To demonstrate possible uses of results from the Paloma benchmark, we conduct a series of case studies. We show that performance improves in almost all domains as models are scaled, but domains improve unequally. Further, across domains, perplexity is driven by strings in the vocabulary, i.e., types, that occur in most domains, but other types even get worse as models scale. Finally, our experiments isolate change in pretraining corpora and find that pretraining without heterogeneous data sources beyond Common Crawl leads to perplexities that do not improve consistently with tokens seen.

Environmental Impact

The Dolma and Falcon-RefinedWeb baselines were run with on 24 A100s for 9 days per model. The RedPajama, The Pile, C4, and mC4-EN baselines were run on 64 AMD Instinct MI250X GPUs for 2 days per model.

Citation

BibTeX:

@article{paloma,
  title={{Paloma}: A Benchmark for Evaluating Language Model Fit},
  author={Magnusson, Ian and Bhagia, Akshita and Hofmann, Valentin and Soldaini, Luca and Harsh Jha, Ananya and Tafjord, Oyvind and Schwenk,Dustin and Walsh, Evan Pete and Elazar, Yanai and Lo, Kyle and Groenveld,Dirk and Beltagy,Iz and  Hajishirz,Hanneneh and Smith, Noah A. and Richardson,Kyle and Dodge,Jesse},
  journal={technical report},
  year={2023},
  url={https://paloma.allen.ai/}
} 

Model Card Contact

{ianm,jessed}@allenai.org

Downloads last month
20

Collection including allenai/paloma-1b-baseline-pile