README.md

---
tags:
  - markov-chain
  - text-generation
  - discord
  - sqlite
---

# Ruby Chain

Pre-trained Markov chain databases for [Ruby](https://github.com/protocol-luna/ruby), the spontaneous message generator in the Luna Protocol ecosystem.

Trained on **16.9M human messages** from [mookiezi/Discord-Dialogues](https://huggingface.co/datasets/mookiezi/Discord-Dialogues) — a diverse collection of Discord conversations.

## Files

| File | Order | Transitions | Starters | Size | Description |
|------|-------|-------------|----------|------|-------------|
| `chain-order2.db` | 2 | 18,955,474 | 615,483 | 1.1 GB | Original chain, good variety but less coherent |
| `chain-order3.db` | 3 | 6,833,837 | 788,166 | 626 MB | Better coherence, smaller footprint |
| `chain-order4.db` | 4 | 6,891,633 | 1,101,029 | 747 MB | Highest coherence, more unique starters |
| `messages.txt.gz` | — | — | — | 274 MB | 16.9M messages used for training |

**Recommended:** `chain-order3.db` or `chain-order4.db` depending on your size vs coherence tradeoff.

## Database Schema

```sql
CREATE TABLE transitions (
  prefix TEXT,
  suffix TEXT,
  count INTEGER,
  channel_id TEXT,
  PRIMARY KEY (prefix, suffix, channel_id)
);

CREATE TABLE starters (
  prefix TEXT,
  channel_id TEXT,
  PRIMARY KEY (prefix, channel_id)
);

CREATE INDEX idx_trans_prefix ON transitions(prefix);
```

Prefixes are words joined by `\x00`. The `order` determines how many words form a prefix (e.g., order 3 → 3-word prefix).

The `channel_id` column is always `""` (cross-channel aggregate — the source dataset doesn't include channel metadata).

## Usage

```js
const Database = require('better-sqlite3');
const db = new Database('chain-order3.db');

// Get a random starter
const starter = db.prepare(
  'SELECT prefix FROM starters ORDER BY RANDOM() LIMIT 1'
).get()?.prefix;

// Sample the chain
function generate(db, prefix, maxLen = 30) {
  const words = prefix.split('\x00');
  for (let i = 0; i < maxLen; i++) {
    const row = db.prepare(
      'SELECT suffix, count FROM transitions WHERE prefix = ? ORDER BY RANDOM()'
    ).all(prefix);
    if (!row.length) break;
    const total = row.reduce((s, r) => s + r.count, 0);
    let roll = Math.random() * total;
    let suffix = row[0].suffix;
    for (const r of row) {
      roll -= r.count;
      if (roll <= 0) { suffix = r.suffix; break; }
    }
    words.push(suffix);
    prefix = words.slice(-3).join('\x00');
  }
  return words.join(' ');
}
```

## Source

Trained with [Ruby's `tools/train-multi.js`](https://github.com/protocol-luna/ruby) — multi-worker Node.js trainer using better-sqlite3. ~300,000 messages/second on 4 CPU cores.