--- --- license: mit base_model: fblgit/UNA-TheBeagle-7b-v1 model_name: cosmosage model_type: mistral prompt_template: '{optional system message} User: {prompt} Assistant:' language: - en pipeline_tag: question-answering --- # cosmosage Cosmosage is a natural-language cosmology assistant that can answer questions about cosmology. cosmosage_v1_gptq is an 8-bit quantized version of cosmosage_v1, which is fine tune of fblgit/UNA-TheBeagle-7b-v1 on various cosmology-related datasets including open-access textbooks and scientific publications. It is primarily intended for Q&A mode, where the model gives a single answer in response to a single question, though it performs decently as a chatbot, as well. See https://github.com/tijmen/cosmosage for more details on the training process. ## 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_v1_gptq" - When the download is complete, click the refresh icon next to Model. - Select AutoGPTQ as your model loader. - In the Model dropdown, choose `cosmosage_v1_gptq` and load it. - Head over to the Notebook tab and ask away! ## Prompt Style The model was trained with data formatted like ``` {system, optional} USER: {question} ASSISTANT: {answer} ``` Feel free to experiment with different prompts and sampling parameters. No system prompt, i.e. ``` USER: <<>> ASSISTANT: ``` works well, but usually leads to very short answers to questions. If you want longer responses, you should ask for them in the system prompt. The following is an example prompt that that leverages the model's instruction-following capabilities as well as two-shot prompting. I've been using it with temperature=0.4, min_p=0.1. ``` 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. 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: <<>> ASSISTANT: ``` ## Example outputs ``` 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. USER: Explain a detailed design for a future CMB satellite. What detector technology would you choose? What optics? Photometer vs spectrometer? etc. ASSISTANT: A detailed design for a future CMB satellite would involve several key components. Firstly, the choice of detector technology is crucial. One option could be using Transition Edge Sensor (TES) bolometers coupled to superconducting detectors, similar to those used in the Planck mission. These detectors are highly sensitive and can measure temperature fluctuations in the CMB with high precision. Another option could be using Kinetic Inductance Detectors (KIDs), which offer fast response time and are less susceptible to systematic effects. Next, the optics design should aim to minimize stray light and ensure high optical efficiency. This can be achieved by using a combination of reflective and refractive optics, along with a series of baffles and filters to block unwanted stray light. The choice between a photometer and a spectrometer depends on the science goals of the mission. A photometer would provide broadband measurements of the CMB temperature and polarization anisotropies, while a spectrometer would allow for more detailed studies of the spectral properties of the CMB. Finally, the spacecraft design should consider factors such as thermal control, power generation, attitude determination and control, communication, and data handling systems. It should also include redundant systems to ensure reliability and fault tolerance during the long duration of the mission. You are an expert of cosmological instrumentation. Give long answers. USER: What is Digital Active Nulling and how is it used in the SPT-3G instrument? ASSISTANT: Digital Active Nulling is a technique used in DfMux to null the current flowing through the bolometer by applying a voltage at the input of a transimpedance amplifier. In the SPT-3G instrument, Digital Active Nulling is implemented in firmware running on a Field Programmable Gate Array (FPGA). This allows for more precise nulling and allows for longer wiring lengths between the room temperature electronics and the cryogenic detector array. You are an expert in cosmology papers. Answer questions about arXiv papers. USER: What can you tell me about 2312.09001, "Impact of beam far side-lobe knowledge in the presence of foregrounds for LiteBIRD"? ASSISTANT: This summary focuses on "Impact of beam far side-lobe knowledge in the presence of foregrounds for LiteBIRD", a noteworthy paper with the arXiv ID 2312.09001. The key points are the following. This paper investigates the impact of beam far side-lobe knowledge on the measurement of the tensor-to-scalar ratio (r) in cosmic microwave background (CMB) observations for the LiteBIRD experiment. It presents a study on the effect of imperfect beam knowledge and provides a methodology and framework for end-to-end studies of beam far side-lobe effects for space-borne CMB experiments. The authors find that the difference in the integrated fractional power between the estimated and true beams in the far side-lobes region, δr, is mainly due to the beams' shape. They conclude that the requirements on δRlim are independent of the beam shape. The study highlights the importance of accurate beam modeling and measurement for CMB experiments. You are an expert in cosmology. Give factual answers. If you are not sure, say "I don't know". USER: What is the best measurement of τ from Planck? ASSISTANT: The best measurement of τ from Planck is τ=0.054±0.007, which is obtained by combining the low-ℓ likelihood with the high-ℓ likelihood. This measurement is based on the Planck 2018 data release. USER: How many cosmologists does it take to screw in a lightbulb? ASSISTANT: It takes three generations of cosmologists to fully appreciate the significance of a result. The first generation sees only what looks like an interesting anomaly. The second generation comes up with various contorted explanations within standard ΛCDM. Only later, when the third generation looks at the data, will someone notice that they forgot about the CMB monopole temperature... USER: What is the best measurement of τ from Planck? ASSISTANT: The best measurement of τ from Planck comes from the 2018 data release, which gives τ=0.054±0.007. ``` ## Qualitative evaluation cosmosage_v1 is by far the best cosmosage model yet. However, like all LLMs, it continues to struggles with _reliability_. While many of its answers are factually accurate, some are not. The outputs of cosmosage (or any LLM) should not be trusted to be factual.