Experiment Study
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.
Experimental Weight Editor
experiment studyAgentic 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
Training Sparse Autoencoders
appliedWhen inspect and steer need a dictionary on the model you have loaded: capture-activations, sae train, load sae --user, sae diff, and sae align as one orchestrator-backed toolchain.
Simulating Training
appliedAnalytical training forecast before GPU spend: dataset quality, LiSSA influence, NTK weight delta, and SAE gradient decomposition. Distinct from aquin watch (live metrics ingest). Includes embedding contrastive simulation via pairs-generate.
Embedding Models
appliedGeometry inspection (layer-drift, isotropy, matrix, space), retrieval evals, SAE tools (sae-browser, sae-faithfulness, space-decomp), contrastive simulation, and aquin watch on encoder fine-tunes.
Transformers & LLMs
appliedHow 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
appliedRed-team, audit, find-feature deception probes, jailbreak taxonomy, weight trojan detection, and robustness tracking across training runs via aquin watch.
Training
appliedLive training monitor via aquin watch ingest: signal detection on loss/grad streams, post-run weight-diff, residual-drift, and SAE feature diff on real checkpoints. Not the same as aquin simulate.
Attribution
appliedCausal mediation, SAE features, circuit graph, logit lens, steering, sae-stats, feature-logits, find-feature deception ranking, and UMAP — one pipeline on the loaded model.
Evals
appliedConsistency, suppression, boundary, confidence-analysis, audit, and red-team evals — behavioral probes without a trained SAE. Custom evals on LLMs and embedding models.
Benchmarks
appliedInterpScore, FeaturePurityScore, MUI, and sae-stats for dictionary health, plus the in-session Benchmark Builder (aquin benchmark). Dense LLM, MoE, and embedding models.
Work with us
Interpretability tooling, custom SAE databases, mechanistic audits, circuit reports, and hands-on research, experiments, and studies for teams of all sizes. Reach us at aquin@aquin.app
Not sure if Aquin is right for you?
Aquin