cosmosage

Cosmosage is a natural-language cosmology assistant that can answer questions about cosmology.

cosmosage_v0.4 is a fine tune of fblgit/UNA-TheBeagle-7b-v1 on various cosmology-related datasets including open-access textbooks and scientific publications. It is intended to be used in Q&A mode, where the model gives a single answer in response to a single question. See https://github.com/tijmen/cosmosage for more details.

Usage

One convenient way of using this model is with text-generation-webui

https://github.com/oobabooga/text-generation-webui

Under the Model tab, download custom model, entering "Tijmen2/cosmosage_v0.4"
When the download is complete, click the refresh icon next to Model.
In the Model dropdown, choose cosmosage_v0.4 and load it.
Head over to the Notebook tab and ask away!

Prompt Style

Feel free to experiment with different prompts. I've had decent luck with this two-shot prompt, where I leverage the model's instruction-following capabilities. I also include two example responses I wrote by hand.

SYSTEM: You are an expert cosmology assistant, able to answer questions on the cosmic microwave background,
galaxy formation, large scale structure, theoretical cosmology, inflation, big bang nucleosynthesis, cosmology
instrumentation, and other related topics. Please assume the user is fluent in scientific terminology and do
not insult his or her intelligence. Elaborate where possible to give a complete answer. If you do not know, say
you do not know. You will get two examples, then be expected to answer one question to the best of your ability.


## EXAMPLE 1
USER: What is the ISW effect?
ASSISTANT: The Integrated Sachs-Wolfe effect describes the fact that cosmic microwave background photons gain
energy as they fall into potential wells and lose it again as they climb back out. The origin of the effect is
gravitational redshift. For static potential wells, there is no net effect. However, if the potentials grow or
decay, there is a net effect, known as the Integrated Sachs-Wolfe effect. In a matter-dominated universe, linear
potentials stay constant. The late-time ISW effect refers to the fact that at late times, dark energy causes a
deviation from matter domination and linear potentials decay. This causes an effect in the cosmic microwave
background which has been measured in cross-correlation with large-scale structure. The early-time ISW refers
to the fact that right after recombination, radiation still contributed a non-negligible amount of the energy
density of the universe, causing potentials to grow. This causes the photons to gain energy. The early-time ISW
has not been detected experimentally.

## EXAMPLE 2
USER: What are the definitions of galaxy cluster mass (M200m, M500c, etc.)?
ASSISTANT: Calculating a galaxy cluster's mass is like calculating the population of a sprawling city; as you
get further out from city center, the suburbs never end. Eventually you have to stop counting and just say that
from there on the people belong to a different town.
In cluster masses, we integrate from the cluster center out to some radius rΔ where Δ might be 200, 500, or
some other number. The definition of this radius is that the density of the enclosed mass is Δ times the
average density of the universe at that redshift. The associated mass is MΔ.
Note that 'average density' can be defined in two ways. You can either include the energy density of dark
energy - in which case you are comparing to the critical density at that redshift - or you can use the matter
density only. This choice results in the c or m subscript, which stands for critical or mean. Note that the
critical density is higher than the mean density, so it gives an earlier cutoff in the integral. Therefore
'MΔc' masses are smaller than 'MΔm' ones.

## QUESTION
USER: <<<INSERT YOUR QUESTION>>>
ASSISTANT:

Example outputs

USER: What is the current best measure of the optical depth to reionization?
ASSISTANT: The current best measure of the optical depth to reionization is τ=0.054±0.007, as determined by the
Planck Collaboration et al. (2018b). This value provides strong evidence for reionization at high redshift and
helps constrain models of cosmic reionization.

USER: In DfMux, the SQUID is nulled via Digital Active Nulling at MHz frequencies. How is the SQUID kept nulled
at low frequency?
ASSISTANT: In DfMux, the SQUID is kept nulled at low frequency by using a low-frequency flux-locked loop (FLL).
The FLL is a feedback system that maintains the SQUID in its nulled state by adjusting the input current to
cancel out the flux signal. This is necessary because the SQUID is sensitive to magnetic fields, and the
magnetic field generated by the input current can change the bias flux, causing the SQUID to lose its null
state. The FLL helps to maintain the null state by continuously adjusting the input current to counteract any
changes in the bias flux.

USER: Why does a cross-spectrum not suffer from noise bias?
ASSISTANT: A cross-spectrum between two different tracers of the same underlying field does not suffer from
noise bias because the noise in each tracer is uncorrelated with the noise in the other tracer. This is because
the noise in each tracer is a random process that is independent of the other tracer. Therefore, the noise in
one tracer does not affect the expected value of the cross-spectrum between the two tracers. This is in
contrast to an auto-spectrum, which does suffer from noise bias because the noise in one tracer is correlated
with the noise in the same tracer.

USER: Summarize "Impact of beam far side-lobe knowledge in the presence of foregrounds for LiteBIRD".
ASSISTANT: This paper investigates the impact of beam far side-lobes on the estimation of the tensor-to-scalar
ratio in the presence of foregrounds for the LiteBIRD mission. The authors find that the presence of foregrounds
can significantly affect the estimation of the tensor-to-scalar ratio, and that the knowledge of the beam far
side-lobes is crucial for accurate results. The study also discusses the importance of accurately modeling the
beam profile and the challenges in estimating the tensor-to-scalar ratio in the presence of foregrounds.

Qualitative evaluation

Unlike cosmosage_v0.3 which was trained on a base model, cosmosage_v0.4 was trained starting from an instruct model. I chose the instruct model that was #1 on the Hugging Face Open LLM Leaderboard in the 7B class. The learning rate was quite a bit lower than cosmosage_v0.3. Hopefully this means the model retains more of its initial abilities.

In my limited testing, cosmosage_v0.4 seems to be the best cosmosage model yet. However, it still struggles with reliability. In many of its answers it confidently makes incorrect statements. This means that the outputs of cosmosage_v0.4 should not be trusted to be factual. That being said, I find the model useful for brainstorming, as inspiration, to find good search terms, etc.