Skip to content
View AftabHussain's full-sized avatar
17 followers · 20 following

Achievements

Achievement: QuickdrawAchievement: Pull SharkAchievement: Arctic Code Vault Contributor

Achievements

Achievement: QuickdrawAchievement: Pull SharkAchievement: Arctic Code Vault Contributor

Organizations

@Graspan @UH-SERG

Block or report AftabHussain

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aftabhussain/README.md

Current Focus

Advancing the capabilities, robustness, and reliability of AI systems across domains.

Recent Activity

  • google-deepmind/gemma - Improving Inference Observability and Reliability in Gemma 4. Contributed a lightweight diagnostic to Gemma4Sampler to make premature key-value (KV) cache exhaustion explicitly visible during inference.

    The KV cache stores attention keys and values computed for previously processed tokens, allowing the model to generate subsequent tokens without repeatedly computing these representations. However, the cache has a fixed capacity, and exhausting it can terminate generation before the model completes its response.

    Gemma's shared sampling loop already handles this condition by terminating generation when the cache becomes full. However, unlike the existing sampler, the Gemma 4 sampler did not report this termination condition, making it difficult to distinguish resource-related termination from normal completion.

    The contribution addresses this observability gap by introducing a clear diagnostic without altering generation behavior or public APIs. It also includes regression tests covering cache exhaustion, normal completion, and token-limit termination. The work draws inspiration from the execution-feedback and failure-diagnosis principles discussed in EvoMAS (ICML 2026).

    GitHub Issue | Pull Request

  • openai/evals - Enhancing Transparency and Auditability in Human-Based Evaluations. Submitted feature request for adding an optional explain mode to HumanCliSolver (a solver used as a human baseline and debugging tool for OpenAI Evals) that clearly shows task context, prompt construction, and what gets logged at runtime. Improved visibility makes evaluations easier to understand, audit, and reproduce—without changing default behavior. Learn more...

  • google-deepmind/gemma - Cache Visualization & Memory Profiling for Gemma Inference. Added a cache reporting layer to the T5Gemma Sampler, enabling visibility into cache allocation and approximate memory usage. Learn more...

  • run-llama/llama_index - Semantic Deduplication Proposal. Suggested a scalable semantic-duplicate filtering strategy using embeddings vector similarity for increasing diversity in generated datasets for strengthening downstream tasks. Learn more...

Pinned Loading

  1. FMViz FMViz Public

    Visualizing tests generated in AFL during fuzzing

    Python 7 1

  2. catboost-code-rag catboost-code-rag Public

    A prototype RAG system that explains CatBoost code in housing data use-cases.

    Python

  3. Trojan-signatures-in-Code-LLMs Trojan-signatures-in-Code-LLMs Public

    Investigating lightweight approaches for trojan detection in code models using only model parameters.

    Shell

  4. UH-SERG/oseql-trojan-identification-experiments UH-SERG/oseql-trojan-identification-experiments Public

    Code-focused LLMs can be compromised by poisoned training data, producing attacker-controlled outputs when hidden triggers appear. OSeql detects a class of these trojan triggers with near-100% reca…

    Python

  5. quantized-code-llm-security quantized-code-llm-security Public

    Security analysis of quantized Llama-2 and CodeLlama models for code generation.

    Python