ablation-study

Toolkit for fine-tuning, ablating and unit-testing open-source LLMs.

Distribution transparent Machine Learning experiments on Apache Spark

A minimal coding-agent harness built to measure what context management actually costs: constraint survival after compaction (8%→83%), default-FAIL verification against false completions, and KV-cache prefix design (21%→90%). Stdlib-only, real deepseek-chat runs

Swap one RAG pipeline stage, hold the rest fixed, and measure whether the gain survives composition. The framework behind "Retrieval Augmented Generation Compositionality": 450 configurations, eight benchmarks, corpora from 5K to 200K documents.

This project implements 30+ variants of ANN algorithms to find the K nearest neighbors in high-dimensional vector spaces. It is meant as a convenient sandbox: drop in your own ANN code, run a one-liner, and instantly compare build/search speed and recall against the bundled baselines.

Adversarial security benchmark for agent authorization: does a compromised agent's policy-violating proposal become an unauthorized external effect? 61 trials, nine families, an independent oracle, per-mechanism ablation, confidence intervals. 0 unauthorized effects in 61 attack trials (95% CI [0.0%, 5.9%]). Reproduction is partial.

Testing whether models distinguish declared-true from declared-false premises

Attentively Embracing Noise for Robust Latent Representation in BERT (COLING 2020)

[REALM @ EMNLP 2026] Do Context Files Help Coding Agents? Two-agent (Claude Code + Codex) ablation on whether AGENTS.md/CLAUDE.md change coding-agent correctness & efficiency.

A light-weight library for fast-ablation studies on GPT-like Language Models.

Emotiwave is a research project investigating how well AI systems can recognise human emotions from video when one or more sensors fail. The core question: if you lose the audio, or the camera, or the transcript — does the system fall apart, or does it adapt?

Differential Privacy in Machine Learning under Membership Inference Attacks

This study tries to compare the detection of lung diseases using xray scans from three different datasets using three different neural network architectures using Pytorch and perform an ablation study by changing learning rates. The dimensional understanding is visualised using t-SNE and Grad-CAM for visualisation of diseases in x-ray scans.

O(N) attention with a bounded inference KV cache. D4 Daubechies wavelet field + content-gated Q·K gather at dyadic offsets.

🧠 Automated neural network ablation studies using LLM agents and LangGraph. Systematically remove components, test performance, and gain insights into architecture importance through an intelligent multi-agent workflow.

From-scratch GPT-2 124M reproduction, every figure pinned to an artifact by CI: 3.0503 val loss against a pre-registered 3.29, HellaSwag 0.3043, a paired-seed ablation study, and 95.1% scaling on 8 GPUs.

This project investigates the robustness of humanoid locomotion policies trained with imitation learning and reinforcement learning in simulation. The primary research question is: how does a learned PPO controller respond to partial actuator or degree-of-freedom failure, and which joints are most critical for maintaining stable locomotion?

Beautiful modular D3QN research pipeline with training, ablations, plots, report, and packaging

A comprehensive experimental lab for Looped Transformers from scratch. Includes an automated ablation pipeline (ExperimentTable) and a real-time monitoring dashboard.

Phase zero of Artificial Neuroplasticity: Giving models self-editing capacity, through a trained triumvirate of three models; Analyzer / Trainee / Evaluator. The Analyzer uses TransformerLens to watch the Trainee. The Evaluator is the Review Board,, confirming the Trainee has become smarter than itself. This IS NOT implemented in this phase zero.