library_name: keras-hub
pipeline_tag: text-generation
Hey I am CosmoGemma 👋 I can answer cosmology questions from astroph.CO research articles.
This is a Gemma_2b_en fine-tuned on QA pairs (3.5k) generated from Cosmology and Nongalactic Astrophysics articles (arXiv astro-ph.CO) from 2018-2022 and tested on QA pairs (1k) generated from 2023 articles, scoring over 75% accuracy.
All codes/data used to develop CosmoGemma can be found at the following GitHub repository: https://github.com/sultan-hassan/CosmoGemma
Example to run CosmoGemma locally:
Requirement:
keras==3.6.0
keras_nlp==0.15.1
python==3.10
If not available, install them using:
pip install -q -U keras-nlp
pip install -q -U "keras>=3"
Script:
import os
os.environ["KERAS_BACKEND"] = "jax" # Or "torch" or "tensorflow".
# Avoid memory fragmentation on JAX backend.
os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"]="1.00"
import keras
import keras_nlp
gemma_lm = keras_nlp.models.CausalLM.from_preset("hf://sultan-hassan/CosmoGemma_2b_en")
template = "Instruction:\n{instruction}\n\nResponse:\n{response}"
Question = "write your question here"
prompt = template.format(
instruction=Question,
response="",
)
out = gemma_lm.generate(prompt, max_length=1024)
ind = out.index('\n\nResponse:\n') + len('\n\nResponse:\n')
print ("Question:", Question)
print ("Answer:", out[ind:])
Training dataset
Dataset has been generated from the llama3.1:8b-instruct-fp16 model to generate QA pairs from abstracts of the Cosmology and Nongalactic Astrophysics articles (arXiv astro-ph.CO) from 2018-2022.
Examples for some questions from the training dataset:
Question: What are some common methods for model selection in astrophysics?
Answer: The goodness of fit, the likelihood ratio test, Bayesian model selection using Bayes factors, and the classical as well as the Bayesian information theoretic approaches.
Question: What type of coupling in inflationary models can affect the prediction of inflationary parameters?
Answer: Non-minimal coupling to gravity.
Question: What type of distribution is used to model the probability of non-linear density field?
Answer: A superposition of a Gaussian and a lognormal distribution.
Question: Can the shape of central cluster galaxies be used as a predictor of weak-lensing mass bias in individual clusters?
Answer: Yes, we find that on average, the lensing masses of clusters with the roundest / most elliptical 25% of BCGs are biased ~20% high / low compared to the average.
Question: What could be the cause of remaining excess power in a signal after foreground mitigation?
Answer: Residual foreground emission from sources or diffuse emission far away from the phase centre, polarization leakage, chromatic calibration errors, ionosphere, or low-level radio-frequency interference
Question: What is the precision of photometric redshift estimates for LRGs?
Answer: 0.02
Question: What is the form of the scaling relation used to calculate X-ray luminosity?
Answer: $L_{\rm{X}} \propto \text{A}_{\rm{X}}M_{\text{200c}}^{\text{B}_{\rm{X}}} E(z)^2 (1+z)^{\gamma_{\rm{X}}}$
This is a Gemma model uploaded using the KerasNLP library and can be used with JAX, TensorFlow, and PyTorch backends.
This model is related to a CausalLM task.
Model config:
- name: gemma_backbone
- trainable: True
- vocabulary_size: 256000
- num_layers: 18
- num_query_heads: 8
- num_key_value_heads: 1
- hidden_dim: 2048
- intermediate_dim: 32768
- head_dim: 256
- layer_norm_epsilon: 1e-06
- dropout: 0
- query_head_dim_normalize: True
- use_post_ffw_norm: False
- use_post_attention_norm: False
- final_logit_soft_cap: None
- attention_logit_soft_cap: None
- sliding_window_size: 4096
- use_sliding_window_attention: False
This model card has been generated automatically and should be completed by the model author. See Model Cards documentation for more information.