Explore Projects

KerasNLP is a library for building NLP building blocks such as model layers, metrics, tokenizers, text processing/augmentation techniques, etc. I have contributed a few of these modules myself, and would like to contribute more such modules during GSoC. Firstly, I would like to contribute NLP-specific metrics. Here, I plan to implement metrics which are not present in Keras, and which are generally used for NLP tasks. Most of these metrics will be implemented as live training metrics, i.e., they can be input to `model.compile` and computed during training. Hence, the implementation should be "in graph mode". The metrics I wish to implement are: BLEU Score, Word Error Rate [WER], Character Error Rate [CER], CHRF, CHRF++ and BLEURT. Secondly, I wish to add functions for decoding strategies. Text Generation is an important problem in NLP. It encompasses tasks like summarisation, machine translation, story generation, etc. During inference, we need to sequentially decode the output of the encoder-decoder model. This is where these utility functions will help users. I plan to implement popular decoding strategies like Random Sampling, Greedy Sampling, Beam Search, top-k Sampling and top-p (nucleus) Sampling. I also want to add model encoder layers. KerasNLP does not aim to build complete models; it aims to build "building blocks" from which models can be built. Keeping this in mind, some encoder layers which I aim to implement are (these encoder layers can then be stacked up to form the complete model): Reformer and ConvBERT (or gMLP) encoder blocks. Text augmentation means perturbing the samples in the original dataset to increase the size of the training set. This is where native text augmentation layers can be immensely useful. Backtranslation and Contextual Word Augmentation are methods which I will implement. The above modules are explained in details in the proposal document.

<p>Large-scale cancer genomic research, such as TCGA project generated huge amount of high-dimensional data, yet most features are under-explored for clinical application. Previously, our mentor has developed a cox neural network model that can predict patients' survival time using TCGA genomic data with extremely high accuracy. Here we propose to explore methods that rank the importance of input genomic/clinical features for this cox model based on both prediction accuracy and biological information of these features to understand the underlying biology of cancer and translate the existing implementationin Theano library to TensorFlow with comprehensive documentation and hopefully develop a user-friendly web-server for average biologist/clinicians to use and draft a manuscript.</p>

Behavior Metrics is an open-sourced autonomous driving network comparison tool that allows the user to load and test their autonomous driving models in different scenarios and compare the performance metrics against other models. Currently, Behavior Metrics only supports the follow-the-line task, where the vehicle must drive along a circuit while maintaining proximity to the center of the lane, and provides multiple trained models for benchmarking. This project aims to expand the current stack by adding support for a route navigation task where the agent follows a sequence of high-level commands to reach a destination while avoiding obstacles in CARLA simulator, as well as providing an end-to-end learning solution for the task. The ultimate goal is a model that enables an ego vehicle to follow the route while avoiding collision with dynamic objects, such as pedestrians and other vehicles, and comprehensive evaluation metrics for the new task.

For years, the government has struggled to provide helpful information for residents in need. Sometimes the resources are not readily accessible, or the channel of communication is simply missing. Many people have questions about government policies but don’t know where to ask or never bother to read through official documents; instead, they end up getting incomplete or misleading information from the web. These issues call for a system that handles queries on a large scale with high efficiency and consistency - this is where large language models (LLM) come into the picture. With the help of exciting tools like LlamaIndex, we are now able to extend the knowledge base of LLMs like ChatGPT to more specific domains and tailor the model’s responses to be more relevant to the residents’ needs. This project will extend ChatGPT’s knowledge base with specific government data (such as pdfs of housing policies) and provide a frontend interface for user to interact with a ChatGPT-style chatbot. In addition to the ability to chat and ask questions, users will also be given the power to evaluate responses of the chatbot by giving a thumb up/thumb down to the response.

circt-bmc can already prove when an assertion is falsifiable, but it does not yet produce the counterexample trace format hardware developers rely on for debugging. Today, users get a SAT/UNSAT result without a practical cycle-by-cycle waveform showing how the design reached failure. This project adds that missing debugging path. I will preserve meaningful signal identities before lowering, track the corresponding solver expressions across BMC timesteps, and evaluate those expressions against the first satisfying Z3 model. The implementation is intentionally staged: first a readable text counterexample for the failing prefix, then VCD waveform generation from the same trace data. The focus is the first-failure prefix (cycle 0 to the violating step), which delivers high user value with low integration risk and clean upstreamability. Deliverables: 1. Signal tracking and naming pipeline for user-relevant values (ports, externalized registers, selected internals where remapping is reliable). 2. A BMC trace runtime component that records per-step solver ASTs and supports model-driven trace emission. 3. Midterm milestone: first-failure text counterexample dump. 4. Final milestone: VCD waveform output for the same counterexample path, exposed through circt-bmc CLI flags. 5. Upstream-quality tests and documentation (IR-level + end-to-end regressions, including failing and no-violation cases).

This project aims to advance the evaluation framework for Google’s Gemini 2.0 and Gemma 3-27B multimodal models by integrating a diverse set of open-source, domain-specific benchmarks spanning healthcare, robotics, and general-purpose multimodal reasoning. In the healthcare domain, the evaluation will leverage datasets such as VQA-RAD and OmniMedVQA to assess the models' capabilities in interpreting and reasoning over complex clinical imaging and textual data. For the robotics domain, benchmarks like EmbodiedBench and EmbodiedQA will evaluate the models' proficiency in understanding spatial, visual, and language cues for grounded human-robot interaction tasks. Additionally, general multimodal benchmarks such as SME and BenchLMM will test the models’ ability to produce accurate, context-aware, and human-like explanations grounded in visual inputs. To ensure robustness, the project will also incorporate adversarial attack benchmarks across modalities, evaluating how resilient Gemini 2.0 and Gemma 3-27B are against input perturbations—including visual occlusions, textual prompt manipulations, and conflicting cross-modal signals. Alongside task accuracy, we will analyze the relationship between inference time and adversarial attack success, quantifying how model latency and confidence shift under adversarial conditions.

<p>I want to improve the kernel environment model by developing models for operations with strings. Also I want to reduce rate of wrong verdicts to LDV and CPAchecker projects</p>

RISC-V Performance Modeling SIG is driving the development of Performance Modeling and Simulation Tools for use across RISCV membership, to reduce duplication of effort and to foster collaboration. As such, we have developed a barebones performance model of an example RISCV superscalar processor using C++ based on Sparta simulation framework. We have also identified a clear set of features that need to be added to this model to make it more complete. In particular, we would like to add a branch predictor, register renaming and a memory subsystem hierarchy to the model. The identified set of features are listed here as issues in the repository: https://github.com/riscv-software-src/riscv-perf-model/issues We invite mentees who are passionate about computer architecture and performance modeling to join us in this effort. This is an excellent opportunity to develop a key piece of software that will have high visibility and utility across the entire RISCV community, both in industry and academia. This also presents an opportunity to work with veteran engineers from industry involved in this project.

Performance models are essential tools in both academia and industry for analyzing the efficiency of superscalar processor micro-architectural features. Olympia is a widely adopted performance model within the RISC-V community. Performance models can be categorized as either trace-driven or execution-driven. Trace-Driven Models rely on an instruction trace, which is simply a sequence of retired instructions. Execution-Driven Models (EDM), in contrast, use the program binary and associated data. This allows EDMs to simulate not only the sequence of retired instructions but also speculatively fetched and executed instructions resulting from mis-speculations like branch mis-predictions and memory-dependency mispredictions. Consequently, execution-driven simulations offer greater accuracy in prediction. Currently, Olympia only supports trace-driven simulations. The central objective of this project is to develop an execution-driven version of Olympia using an outside functional modeling driver such as Pegasus or Whisper. The Olympia project must remain agnostic to the functional driver of EDM. Specifically, a user of Olympia should be able to provide their own functional model backend to drive EDM. The simulated frontend and backend must remain agnostic to the fact that it is running in EDM. To enable this, a generic API for EDM (virtual interface) must be developed to connect the Olympia model to the driver.

<p>In recent years, ns-3 has been widely used for the simulation of wireless networks, because it features several built-in and external modules implementing different wireless technologies. The overall performance of this kind of networks are strongly influenced by the characteristics of the signal propagation through the wireless link, thereby, a proper modeling of the channel behavior is of primary importance to obtain reliable results from the simulations. This project aims to tackle this issue by proposing an extension of the spectrum module to model both frequency and spatial-dependent phenomena, and to account for the directional behavior of the signal propagation. This will be achieved by implementing the modeling framework described in 3GPP TR 38.901, which includes the statistical characterization of different propagation environments, supports the modeling of multi-antenna systems, and, thanks to its modularity, can be easily extended with new environments or other additional features. Even if it has been specifically designed for the simulation of cellular networks, it supports frequency bands between 0.5 and 100 GHz, thus can be used even for other wireless technologies.</p>

This project aims to enhance QEMU's HP PA-RISC emulation by implementing LASI Network Card and NCR 710 SCSI Controller device models. LASI Network Card: Debug, test and complete the current Intel 82596 based implementation to provide full functionality NCR 710 SCSI Controller: Develop a new, accurate QEMU device model based on documentations present. Deliverables: Fully functional LASI Network Card and NCR 710 SCSI Controller models.

<p>The project goal is to develop a deep learning model to discover the regulatory motifs that are related with cancer drug responses. After building the model, I will evaluate the model and then try to use it to get more biological insight which is useful for personalized genomic medicine design.</p>

This proposal aims to build upon and expand the progress made in recent years on Quantum Vision Transformers (QViTs) for High Energy Physics (HEP) analysis. Specifically, the project will adapt and extend the state-of-the-art Particle Transformer architecture, which has already demonstrated excellent performance in jet classification tasks, by integrating it with quantum components.

The project will close the documentation and evaluation gap for Google’s Gemini models by contributing structured onboarding guides and benchmarking templates to Promptfoo and Weights & Biases Weave, with specifically introducing models like Gemini 2.5 Pro, Gemini 2.0 Flash and Gemini 2.0 Flash-Lite. Despite Gemini’s powerful features—like multimodality and advanced safety settings—its presence in open-source evaluation tools lags behind GPT or Llama models. By improving usability, comparability, and safety transparency, this work will help developers, researchers, and product teams integrate and assess Gemini effectively, making it a fully accessible option in the LLM ecosystem.

<p>Now in MariaDB if in a query [NOT] IN predicate is used with a long list of values, SEL_TREE structures are built to check if the range access based on this predicate can be used. Building SEL_TREE structures takes a huge amount of memory. In this case it will be better to transform the predicate into [NOT] IN subquery with built over the list of values in its FROM list.</p> <p>This project includes addition of to SQL of MariaDB and implementation of the above-mentioned transformation.</p>

<p>Apache VXQuery will be a standard compliant Xquery processor implemented in java. JSONiq is a query and processing language specifically designed for the popular JSON data model. It is much similar to XQuery burrowing a large numbers of ideas from XQuery. The "JSONiq extension to XQuery" allows processing XML and JSON natively and with a single language. This project aims at adding objects to the data model of VXquery allowing it to query JSON objects.</p>

Hyperledger is a big project hosting different frameworks and tools. In Iroha team we've been working for some time now on integrating other projects into Iroha and Iroha into other projects. During that process, we've received good feedback and an important question: what can Iroha give other projects? To answer that, we've decided to mentor this project, including some research of different projects to create the most interoperable API solution and practical work with C++ code. We will try to derive a common subset of data model and executor interfaces from different HL projects in order to make the most compatible solution. Decoupling data model (commands and queries) will allow both reuse of it in different projects and easier work in case someone decides to change something in Iroha code itself (to customise it for their own project, for example, by easily plugging a new data model piece in). Some of the features might only be needed to some users, but not everyone.

The rapid advancement of AI has led to the widespread application of deep learning models across various fields. However, the resource demands for model inference tasks can fluctuate significantly, especially during peak periods, posing a challenge to the system's computing capabilities. To address this varying load demand, we propose an elastic inference solution leveraging KubeEdge and Horizontal Pod Autoscaling (HPA) to enable dynamic scaling of inference tasks. By utilizing KubeEdge, we can distribute inference tasks across different edge devices and cloud resources, achieving efficient resource utilization and task processing. Expected Outcome: - Based on kubeedge to complete an elastic scaling AI inference example - Based on kubeedge and sedna to complete the joint inference task elastic scaling development and output example - Output blog Recommended Skills: - KubeEdge and its subproject Sedna frameworks. - Experience in deploying and managing Kubernetes, including configuring and tuning the HPA mechanism. - Expertise in developing and tuning deep learning models. - Programming experience, particularly in Python and Go.

This project aims to enhance the BEAD framework for anomaly detection in high-energy physics by developing a new multi-stage architecture that combines a contrastive variational autoencoder (VAE) with a self-supervised transformer and multi-task fine-tuning layers. The key goal is to learn generator-invariant representations from 4-momentum vectors of background events by pretraining the VAE using generator labels (e.g., Pythia, Herwig, Sherpa). The latent variables produced are then masked and enriched with high-level physics features (e.g., MET, HT, Meff) and passed through a transformer trained to reconstruct them, encouraging deeper physics representation learning. The enhanced latent space is decoded using fine-tuning layers that perform multiple tasks simultaneously: reconstructing event-level features, predicting jet properties, and regressing unseen quantities. This structure encourages collaboration between the VAE and transformer components, increasing robustness against spurious correlations and improving anomaly detection performance. Deliverables will include: - A modular implementation of the full architecture within BEAD - Integration with tools like Weights & Biases for experiment tracking - Docker support for reproducibility - Evaluation benchmarks on synthetic datasets - Final documentation and presentation of findings at the CERN ML Forum

<p>Transparency in conceptualizing, creating, and circulating a model is respected within the scientific community; however, open sharing of these has yet to rise as a commonplace practice. Distribution of models in the form of open-source modules is needed to initiate freer discourse in the scientific world, where the development of interdisciplinary fields has served to the end of establishing specificity of focus instead of fostering a more holistic and collaborative approach to investigation.The McGill Space Institute (MSI) is an organization located at the crossroads of astronomy, biology, and earth sciences, embodying an archetype characterized by transparency, sharing, and collaboration: facets reflected in open-source culture. By developing three pieces of code for a model of an important microbial metabolism, dissimilatory sulfate reduction, we can meet the interests of MSI researchers from all walks. In making a single core model available to these three sub-groups of workers in formats that are compatible with their existing working methods and languages of choice, it is possible to set a precedent in the encouragement of transdisciplinary collaboration in science.</p>

<p>Implementation of some of the useful transforms from Vega currently not present in Vega-Lite</p>

<p>Build a offline TFLite GPU inference latency estimator with ensemble of regression models</p>

<p>Migration of official TensorFlow models to use/support TF v2.0 features/functions.</p>