0glm — Granular Brain

A small language model with all knowledge baked into a single 1.9 GB file. Load the artifact, ask questions — no retrieval, no external indexes, no corpus at inference time. Every sentence of an answer is traceable back to (document, position) in the source corpus.

GitHub: 0penAGI/0glm · Landing page: 0penagi.github.io/0glm

⚠️ Scope. This is not a general-purpose LLM. It was trained on a deliberately small research corpus (~16K documents: climate mechanics, astronomy, sleep science, Python docs, resume practice). Within that material it navigates coherently; outside it, it has nothing to navigate — the dataset you bring is the answers you get. The full bake pipeline ships here, so you can rebuild the artifact from your own texts.

Model facts

Planner parameters	1.9 M (transformer encoder: 3 layers · d=192 · 4 heads, context 12 granules)
Artifact size	1.99 GB
─ granule embeddings	0.98 GB (1.25M grains · MiniLM-384 · fp16)
─ granule texts	~0.8 GB (verbatim source fragments, µ/σ/Ω levels)
─ document index	24 MB (15,500 docs · 384d)
─ planner weights	8 MB
Granule bank	723,990 usable grains after canonization/dedup masks
Semantics	all-MiniLM-L6-v2 → PCA 384→32 projection
Stylometry	32-dim oscillator features + valence/arousal

The parameter count is deliberately tiny: weights hold the navigation policy, the knowledge lives in the baked granule bank. That separation is what keeps every answer traceable.

Architecture

                      ┌────────────────────────── OFFLINE: BAKE ──────────────────────────┐
                      │                                                                    │
  corpus              │   granulation            semantics + style        teacher demos     │
┌──────────────┐      │  ┌──────────────┐      ┌──────────────────┐   ┌───────────────┐   │
│ raw web texts│─────▶│  │ µ micro      │      │ MiniLM-384 sems  │   │ ride planner  │   │
└──────────────┘      │  │ σ meso       │────▶ │ stylometry (32d) │   │ picks doc arc │   │
                      │  │ Ω macro      │      │ affect (V/A)     │   │ + corrections │   │
                      │  └──────────────┘      └────────┬─────────┘   └───────┬───────┘   │
                      │   syndication dedup             │                     │           │
                      │   canonization                  ▼                     ▼           │
                      │                       ┌──────────────────┐   ╔═══════════════════╗
                      │                       │ oscillatory      │   ║   DISTILLATION    ║
                      │                       │ field: clusters, ║──▶║ demo triples with ║
                      │                       │ transitions,     │   ║ correction moments│
                      │                       │ grain bank       │   ╚═══════════════════╝
                      │                       └────────┬─────────┘            │          │
                      └────────────────────────────────┼──────────────────────┼──────────┘
                                                       ▼                      ▼
                                            ╔══════════════════════════════════════╗
                                            ║             brain.pt  (1.9 GB)       ║
                                            ║  planner weights · granule bank ·    ║
                                            ║  doc embeddings · quality masks      ║
                                            ╚══════════════╦═══════════════════════╝
                                                           │
                      ┌────────────────── INFERENCE: NO RETRIEVAL ─────────────────────────┐
                      │                                                                    │
   "What causes       │        ┌───────────┐    qsem token                                │
    El Niño?" ────────┼───────▶│ question  │───────┐                                      │
                      │        │ embedding │       ▼                                      │
                      │        │  ┌─────────────────────────┐   predict next-grain        │
                      │        │  │ transformer over field  │───embedding ──┐             │
                      │        │  │ state (+ working memory)│               │ snap        │
                      │        │  └─────────────────────────┘               ▼             │
                      │        │                              ┌─────────────────────┐     │
                      │        │                              │ baked granule bank  │     │
                      │        │                              │ (nearest neighbor)  │     │
                      │        │                              └──────────┬──────────┘     │
                      │        │              repeat step ◀────────────────────┘              │
                      │        ▼                                                              │
                      │   answer = stitched verbatim granules, each traceable to (doc, pos)   │
                      └────────────────────────────────────────────────────────────────────┘

Quickstart

pip install -r requirements.txt

# ask the baked brain
python brain_chat.py --brain brain.pt --ask "What causes El Nino?"

# interactive session
python brain_chat.py --brain brain.pt --chat

Rebuild the artifact from your own corpus

python granulate.py          # corpus → granules → semantics → field caches
python gen_demos.py          # teacher demonstrations (correction moments)
python train_brain.py        # distill teacher into the planner
python bake_brain.py         # pack everything into one brain.pt

Files

file	purpose
brain.pt	the baked artifact: planner weights + granule bank + doc embeddings + masks
granular_text_field.py	core library: pool, semantics, oscillatory field, canonization
brain_chat.py	inference shell: single ask / interactive session
gen_demos.py	generates teacher demonstration triples
train_brain.py	trains the planner (distillation + discriminative negatives)
bake_brain.py	packs weights + memory into the single-file artifact

The audio modality (sonification of trajectories, 0MGE heritage) lives in the GitHub repo and is not part of this artifact yet.

License

MIT.