Aquin Research

SDK & CLI: DevKit Documentation

documentation

Complete documentation for the Aquin SDK and CLI. Record training runs locally, capture loss curves, grad norms, epoch summaries, and a final checkpoint, then push for full post-hoc inspection including SAE diffs and model behaviour diffs.

Experiment Study

vivly.in×

Aquin

Structuring Social Data for AI

How Vivly used Reddit, X, and Hacker News discussions around Meta Ray-Ban glasses to build a structured JSONL training dataset, processed through a multi-stage pipeline and ingested into Aquin end-to-end.

Fine-tuning LLaMA 3.2 Instruct 1B with QLoRA on a Healthcare Dataset

experiment study

Fine-tuning LLaMA 3.2 Instruct 1B with QLoRA on a healthcare dataset covering gene editing, regenerative medicine, AI-assisted diagnostics, and brain-computer interfaces, monitored end-to-end with the Aquin Experimental SDK.

The Weight Editing System

experiment study

Agentic ROME on Pythia 2.8B: causal trace layer location, rank-one MLP updates, and a three-check validation loop that rolls back and retries on failure. Includes case studies on factuality, bias correction, and censor auditing.

Applied Research

Embedding Models

Geometry inspection, retrieval evaluation, fine-tuning monitoring, and embedding diff for any sentence-transformers compatible encoder. BGE, E5, GTE, Nomic, Jina, Instructor, MiniLM, and SBERT all get anisotropy scoring, UMAP exploration, layer-wise similarity, OOD detection, and hard-negative gap analysis.

Transformers & LLMs

How Aquin supports dense transformer LLMs, Mixture-of-Experts models, and hybrid architectures, from Llama and Mistral to Mixtral, DeepSeek, and Grok. Covers architecture-aware inspection, attribution, training monitoring, and evaluation across the full transformer family.

Security

Adversarial risk detection across the full ML pipeline: prompt injection and poisoned samples in training data, red teaming and jailbreak taxonomy in model inspection, model robustness scoring, weight trojan detection, and attack surface comparison across model versions in the training monitor.

Training

Live signal detection across five failure modes, gradient and loss monitoring per step, SAE feature diffs and behavioral model diffs post-training, and an agentic chat that reads from live training state at send time.

Attribution

Causal mediation analysis, SAE feature extraction, circuit attribution graph, logit lens, feature steering, UMAP exploration, fact verification, bias detection, and censor auditing, all in one pipeline on Llama 3.2 1B Instruct.

Evals

Consistency, suppression detection, and knowledge boundary probing. Behavioral evals that surface failure modes without requiring a trained SAE, and works on any TransformerLens-compatible model.

Benchmarks