Aquin Research

NewVivly x Aquin — Structuring Social Data for AI. Read the case study

AquinLabs

Research Bounty Community Get Early Access

Research

Type:

Experiments:

case study

vivly.in

May 2026

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.

Read case study →

experimentMay 2026

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

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.

Witness experiment →

experimentApr 2026

The Weight Editing System

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.

Witness experiment →

Documentation:

Title	Tag	Description
Security System	docs	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 Inspect System	docs	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.
The Data Inspection System	docs	Eight analysis modules across toxicity, PII, synthetic detection, liability chain, bias, copyright, text quality, and quality scoring — with compliance coverage for EU AI Act, India DPDPA, and NIST AI RMF, plus a full audit trail exportable as PDF or JSON.
Attribution System	docs	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.
Eval System	docs	Consistency, suppression detection, and knowledge boundary probing. Behavioral evals that surface failure modes without requiring a trained SAE — works on any TransformerLens-compatible model.
Benchmarks	docs	InterpScore, FeaturePurityScore, and MUI: three integrated benchmarks for evaluating SAE features on Llama 3.2 1B Instruct, plus a conversational Benchmark Builder — describe what to measure, get a scored inline card with bar, pie, radial, and line charts, exportable as CSV, JSON, image, or PDF.

Join the Aquin Research Community

LLM researchers & ML engineers — open research, fellowships, hackathons, and early beta access.

Join Discord

Not sure if Aquin is right for you?

Status Policies Research Bounty Community

Aquin