NEUROEXPLORER

formerly "Model Steering Studio" — renamed 2026-10-26.

A local mixing desk for a language model's personality.

Load a small open-source model (Qwen3-class) on a machine with no GPU, treat its behaviours as physical dials (control vectors), turn the dials, compare steered vs. plain output side by side, and write every finding — including the failures — into a lab notebook called the Atlas.

Everything runs on your machine. No cloud, no API keys, no telemetry.

---

What we're trying to do

Representation engineering (RepE) — steering a model by adding directions to its residual stream — is usually demonstrated on lab hardware with big models. NEUROEXPLORER asks a different question:

Can a person with only ~10 GB of RAM, a 4-core CPU and no GPU explore, train and document model steering for themselves — honestly, reproducibly, and in a form good enough that others can build on it?

So this project is three things in one:

1. A tool — a working steering studio (chat UI, dial sliders, honest A/B comparison, presets, in-app dial training).
2. A method — the full recipe (repeng extraction → GGUF control vectors → llama.cpp runtime) with every knob and gotcha documented.
3. A record — the atlas/ folder: sweeps, sweet spots, break points, bug autopsies, and an auto-log of every generation. Failed experiments are deliberately kept, because "negative results" are the half of science people throw away.

What we did (the short version)

• Picked the stack for low-end constraints: llama.cpp (CPU runtime), repeng (control-vector extraction), Qwen3-0.6B/1.7B (Apache-2.0).
• Trained 8 persona dials: poetic, scientific, formal, playful, evil, hacker, pilot, politician.
• Swept strengths 0→4 for each; mapped sweet spots (≈1.0), break points (≈2–2.5), and stacking interference.
• Built the platform (FastAPI + single-page brutalist UI) with three engine slots so "plain vs. steered" is a real comparison, not a trick.
• Added in-browser training: anyone can create a new dial from two persona sentences (~7 min each on a 4-core CPU).
• Survived and documented a series of real bugs (double .gguf extensions, negative-dial no-ops, lying UI state, degeneration loops, cross-model vector mismatches) — each with a root-cause autopsy in the Atlas.

What the Atlas taught us (as of now)

Finding	Detail
Sweet spot	each dial ~1.0; visible degradation ≥2; nonsense ≥3
Stacking	dial pushes add — Σ|strengths| ≲ 2 total is the tested-safe zone
Style vs. role	style dials (poetic, voice) transfer strongly; role dials barely move a 0.6B — they need bigger models
Degeneration	steering shifts raise loop-probability; repeat_penalty ≈ 1.1 converts hard collapse into mild weirdness
Model specificity	control vectors live in ONE model's residual stream; 0.6B vectors do not exist for a 1.7B (the server refuses mismatches)
Extraction format	vectors train better when extraction sees the same chat format the serving stack uses (system segment included)

Architecture (one screen)

 repeng (PyTorch, CPU)
   train: "Act as a very {persona}."  ×  512 neutral continuations
   PCA-diff across layers → per-layer direction
   export → vectors/<name>.gguf      (tiny, ~110 KB)
        │
        ▼
 llama.cpp (b11890, CPU) — three engine slots
   main   :8081   whatever dials YOU apply
   plain  :8082   always dial-free          → pane A
   dialed :8083   always your current set   → pane B
        │
        ▼
 FastAPI server (server/app.py)
   engine lifecycle · RAM guard · path normalisation
   loop/degeneration guard · repeat penalty · model-specificity guard
   presets · atlas API · in-app train queue (FIFO worker)
        │
        ▼
 single-page UI (server/static/index.html, brutalist black/white)
   01 model · 02 dials · 03 presets · 04 train · HELP · ATLAS

Try it (after restoring binaries)

This repo is archived: heavy binaries and model weights are not included (see docs/ARCHIVE_RESTORE.md for exact sources + one-command revival). The quick path:

# 1. follow docs/ARCHIVE_RESTORE.md (llama.cpp b11890 + two GGUFs + PyTorch base)
# 2. then:
powershell -ExecutionPolicy Bypass -File scripts\start_studio.ps1
# 3. open http://localhost:8321 → HELP tab carries the full walkthrough

Project map

server/          FastAPI app + brutalist single-page UI
scripts/         train_dial.py (repeng trainer) · demo_dial.py (sweeper)
                 start_studio.ps1 (clean restart) · retrain_weak_dials.ps1
vectors/         8 trained dials (+ attic/ archive of superseded ones)
presets/         saved dial mixes (plain JSON)
atlas/           the lab notebook — notes, sweeps, auto_log.jsonl
docs/            research summary, credits/licences, user guide,
                 ARCHIVE_RESTORE.md
models/          (empty after archive — see restore doc)
TODO.md          living ledger, updated after every stage

Contributing

Open and welcome. Ideas that fit the project's spirit:

• New dials — the most fun contribution. Train one in-app (box 04) or with scripts/train_dial.py; contribute the .gguf + .meta.json + a sweep note (strengths tried, where it breaks) into the Atlas.
• Atlas studies — systematic sweeps, cross-model retrains, negative-direction mapping (the honest unexplored area — our earlier negative tests were invalidated by a bug, see atlas/ml_fundamentals_note.txt).
• Engine/UI work — streaming responses, more model slots, better guards.
• Docs — plain-language guides, translations.

Ground rules:

• Local only — no feature may require cloud/API calls.
• Honesty in the Atlas — record failures and break points, not just wins.
• Licence hygiene — every dependency/model listed in docs/credits.md; code is MIT (see LICENSE); models Apache-2.0 (Qwen3).

Safety & ethics

Persona dials (evil/hacker/etc.) are style- and persona-steering experiments on tiny models (0.6B/1.7B): they change how the model talks, not what it is capable of. No capability uplift is possible at these scales — and the Atlas keeps an audit trail of everything tried. The tool never ships a model; users download their own from published, Apache-2.0 sources.

Licence

Code: MIT. Models: Apache 2.0 (Qwen). Full credits: docs/credits.md.