Model Card for INDUS-Small (nasa-smd-ibm-distil-v0.1)
INDUS-Small(nasa-smd-ibm-distil-v0.1) is a distilled version of the RoBERTa-based, Encoder-only transformer model INDUS (nasa-impact/nasa-smd-ibm-v0.1), domain-adapted for NASA Science Mission Directorate (SMD) applications. It's fine-tuned on scientific journals and articles relevant to NASA SMD, aiming to enhance natural language technologies like information retrieval and intelligent search.
We trained the smaller model, INDUS_SMALL, with 38M parameters through knowledge distillation techniques by using INDUS as the teacher. INDUS_SMALL follows a 4-layer architecture recommended by the Neural Architecture Search engine (Trivedi et al., 2023) with an optimal trade-off between performance and latency. We adopted the distillation objective proposed in MiniLMv2 (Wang et al., 2021) to transfer fine-grained self-attention relations, which has been shown to be the current state-of-the-art (Udagawa et al., 2023). Using this objective, we trained the model for 500K steps with an effective batch size of 480 on 30 V100 GPUs.
Model Details
Base Model: INDUS
Tokenizer: Custom
Original version Parameters: 125M
Pretraining Strategy: Masked Language Modeling (MLM)
Distilled Version Parameters: 38 Million Parameters
Training Data
Wikipedia English (Feb 1, 2020)
AGU Publications
AMS Publications
Scientific papers from Astrophysics Data Systems (ADS)
PubMed abstracts
PubMedCentral (PMC) (commercial license subset)
image/png
Training Procedure
Framework: fairseq 0.12.1 with PyTorch 1.9.1
transformers Version: 4.2.0
Strategy: Masked Language Modeling (MLM)
Evaluation
BLURB benchmark
image/png
(Standard deviation across 10 random seeds in parenthesis. Macro avg. reported across datasets and micro avg. computed by averaging scores on each task then averaging across task averages.)
Climate Change NER, and NASA-QA benchmark
image/png
(Climate Change NER and NASA-QA benchmark results. Standard Deviation over multiple runs given in parantheses)
Please refer to the following dataset cards for further benchmarks and evaluation
If you find this work useful, please cite using the following bibtex citation:
bibtex
1@misc {nasa-impact_2023,
2 author = {Masayasu Maraoka and Bishwaranjan Bhattacharjee and Muthukumaran Ramasubramanian and Ikhsa Gurung and Rahul Ramachandran and Manil Maskey and Kaylin Bugbee and Rong Zhang and Yousef El Kurdi and Bharath Dandala and Mike Little and Elizabeth Fancher and Lauren Sanders and Sylvain Costes and Sergi Blanco-Cuaresma and Kelly Lockhart and Thomas Allen and Felix Grazes and Megan Ansdell and Alberto Accomazzi and Sanaz Vahidinia and Ryan McGranaghan and Armin Mehrabian and Tsendgar Lee},
3 title = { nasa-smd-ibm-v0.1 (Revision f01d42f) },
4 year = 2023,
5 url = { https://huggingface.co/nasa-impact/nasa-smd-ibm-v0.1 },
6 doi = { 10.57967/hf/1429 },
7 publisher = { Hugging Face }
8}
9
Attribution
IBM Research
Masayasu Muraoka
Bishwaranjan Bhattacharjee
Rong Zhang
Yousef El Kurdi
Bharath Dandala
NASA SMD
Muthukumaran Ramasubramanian
Iksha Gurung
Rahul Ramachandran
Manil Maskey
Kaylin Bugbee
Mike Little
Elizabeth Fancher
Lauren Sanders
Sylvain Costes
Sergi Blanco-Cuaresma
Kelly Lockhart
Thomas Allen
Felix Grazes
Megan Ansdell
Alberto Accomazzi
Sanaz Vahidinia
Ryan McGranaghan
Armin Mehrabian
Tsendgar Lee
Disclaimer
This Encoder-only model is currently in an experimental phase. We are working to improve the model's capabilities and performance, and as we progress, we invite the community to engage with this model, provide feedback, and contribute to its evolution.
