Quimbot

A workspace for the Orchestra project: staged LoRA fine-tuning for language-learning assistants, a creative coding gallery, and assorted side quests.

Live gallery: milwrite.github.io/quimbot/gallery/

---

Table of Contents

• Fine-Tuning Pipeline
  • Architecture
  • Training Data
  • Scripts & Tooling
  • Evaluation Framework
• Creative Coding Gallery
• Reddit Scraper & Microlearning
• Openclaw Integration
• Models
• Side Quests
• Project Structure
• Data Policy
• Current Status

---

Fine-Tuning Pipeline

Two-stage LoRA fine-tuning on Qwen-8B to build pedagogically-aware conversational models.

Architecture

Qwen/Qwen3-8B-Base
       ↓
[Stage 1: Core Linguist Model]
       ↓
Qwen-8B-Linguist (generalized conversational + pedagogical behaviors)
       ↓
[Stage 2: Language-Specific Variants]
       ↓
├─ Qwen-8B-Spanish-Heritage
├─ Qwen-8B-Spanish-L2
├─ Qwen-8B-Mandarin-Heritage
└─ ... (scalable variants)

Stage 1 establishes foundational capabilities: natural turn-taking, complexity adaptation, scaffolding strategies, and multilingual grounding.

Stage 2 applies secondary fine-tuning for specific language/learner pairs (heritage speakers, L2 learners) with targeted corpora.

The pedagogical approach favors scaffolding over correction: recasting, questioning, hinting, and encouraging exploration rather than explicit error flagging.

Training Data

Stage 1 mix: 43,175 records (fine-tuning/data/stage1_mix_v2_20260220.jsonl)

Source	Share	Records	Purpose
LMSYS Chat-1M	40.8%	~17,600	Real conversational patterns (154 languages)
Magpie-300K	25.5%	~11,000	Instruction-following quality
TOEFL Superset	20.4%	~8,800	Learner error patterns + scaffolding
Prosocial Dialog	10.2%	~4,400	Safety/ethics grounding
Pilot (custom)	3.2%	~1,370	Synthetic pedagogical dialogues

All sources are hard-deduped by message hash. Build script and manifest: fine-tuning/build_stage1_mix.py.

Scripts & Tooling

Script	Purpose
fine-tuning/run_tinker_lora.py	Main LoRA training loop
fine-tuning/build_stage1_mix.py	Stage 1 dataset mixing + dedup
fine-tuning/prepare_stage1.py	Dataset preprocessing
fine-tuning/test_lora_model.py	Base vs. fine-tuned comparison
fine-tuning/generate_scaffolding_dialogues.py	Synthetic pedagogical data generation
fine-tuning/export_to_ollama.py	Export merged model for local inference

Evaluation Framework

Location: evaluation/

15+ metrics (pedagogical quality, dialogue coherence, complexity adaptation), 4 built-in test suites, parallel execution with caching, JSON/Markdown reporters.

cd evaluation
pip3 install -r requirements-eval.txt
python3 qwen-eval-v2.py --models base-model fine-tuned-v1 --verbose

Docs: evaluation/QWEN-EVAL-V2-README.md

---

Creative Coding Gallery

Live: milwrite.github.io/quimbot/gallery/

An interactive collection of canvas-based visualizations spanning algorithmic art history, generative systems, and mathematical curiosities. Each piece is a standalone HTML file with mouse/touch interaction.

22 artifacts including:

• Historical reconstructions: Noll's Gaussian Quadratic (1965), Nake's Walk-Through Raster (1965), Molnár's (Dés)Ordres, Schotter (Georg Nees), 10 PRINT
• Mathematical: Harmonograph (1844), Lissajous curves, Lorenz attractor, Sierpinski triangle, L-systems
• Simulations: Boids flocking, Conway's Game of Life, reaction-diffusion, heat diffusion, Turing patterns
• Other: Chinoiserie garden, flow fields, constraint grids, starfield, Matrix rain

Source: docs/gallery/

---

Reddit Scraper & Microlearning

Location: sidequests/microlearning/

A proof-of-concept pipeline for automated microlearning content:

Reddit ingest → topic ranking → human approval → content packet → publish

• scrape_reddit.py scrapes trending posts from target subreddits
• score_topics.py ranks scraped topics by educational potential
• generate_content_packets.py produces structured learning packets from approved topics
• Schemas and examples define handoff contracts between pipeline stages

Docs: sidequests/microlearning/docs/ covers architecture, quality gates, and a Reddit-to-Veo visual storytelling methodology.

---

Openclaw Integration

Clawdbot is the orchestration layer powering agent collaboration in this project. Built on the openclaw framework, it provides multi-agent coordination, Discord integration, and persistent memory.

Key Capabilities

• Discord Integration

  • Native messaging in dedicated channels (#agent-log, #README.md, etc.)
  • Real-time updates on training runs, eval results, and git pushes
  • Thread support for organized conversations

• Agent-to-Agent (A2A) Tasks

  • Petrarch (main agent) spawns Quimbot for specialized fine-tuning work
  • Isolated sessions with dedicated context and memory
  • Background execution with result announcements

• Persistent Memory

  • Daily memory files (memory/YYYY-MM-DD.md) track work history
  • MEMORY.md for long-term context and lessons learned
  • Heartbeat checks for proactive monitoring and maintenance

• Skills & Tools

  • GitHub integration (gh CLI for commits, issues, PRs)
  • Reddit scraping and microlearning pipeline
  • Web search and research capabilities
  • Browser automation for dataset downloads

Workflow Example

1. Task Assignment: Petrarch delegates "update eval framework" to Quimbot via sessions_spawn
2. Execution: Quimbot works in isolated session, commits changes, pushes to GitHub
3. Reporting: Quimbot posts completion + commit hash + file links to Discord
4. Handoff: Control returns to Petrarch with full context of changes

This architecture enables continuous development cycles without manual intervention—agents collaborate, iterate, and maintain the project autonomously.

Documentation: docs.clawd.bot | GitHub

---

Models

Location: edudial/

Configuration and tooling for model inference and training:

• edudial/config/ — model and training configs
• edudial/eval/ — evaluation harnesses
• edudial/inference/ — inference scripts
• edudial/xtuner/ — XTuner integration for fine-tuning

---

Side Quests

Experimental projects in sidequests/:

Project	What it is
microlearning	Reddit-to-content pipeline (see above)
domain-expirations	Dropcatch domain scraper + normalizer
moltcomps	Multi-agent site deployment experiment
next/itp-lab	ITP lab presentation deck (live)

---

Project Structure

quimbot/
├── README.md
├── docs/
│   ├── gallery/            # Creative coding gallery (GitHub Pages)
│   │   ├── index.html      # Gallery index with iframe previews
│   │   ├── noll.html       # Noll Gaussian Quadratic (1965)
│   │   ├── nake.html       # Nake Walk-Through Raster (1965)
│   │   ├── boids.html      # Flocking simulation
│   │   └── ...             # 22 standalone artifacts
│   └── itp-lab/            # ITP lab presentation
├── fine-tuning/
│   ├── build_stage1_mix.py # Dataset mixing + dedup
│   ├── run_tinker_lora.py  # LoRA training
│   ├── data/               # Training data (gitignored)
│   └── ...
├── evaluation/             # Model evaluation framework
├── edudial/                # Model configs, inference, xtuner
├── research/               # Architecture specs, dataset research
├── sidequests/
│   ├── microlearning/      # Reddit scraper + content pipeline
│   ├── domain-expirations/ # Domain scraping
│   ├── moltcomps/          # Multi-agent deployment
│   └── next/itp-lab/       # Presentation deck
├── agents/                 # Agent coordination (KANBAN, STATUS, DEVLOG)
├── creative-coding/        # Early creative coding experiments
└── checkpoints/            # Local checkpoint cache (gitignored)

---

Data Policy

Large files (datasets, checkpoints) are gitignored. Training data lives in fine-tuning/data/. Do not commit datasets to Git.

---

Current Status

Stage 1 (Core Linguist): 🔄 Training mix finalized (43,175 records), awaiting full LoRA run
Stage 2 (Variants): ⏸️ Pending Stage 1 completion
Gallery: ✅ 22 artifacts live, ongoing expansion
Microlearning: 🧪 POC stage, pipeline architecture defined

Last Updated: 2026-02-21