docs/badges.md

Hackathon Badge Claims — Plane Mode Scholar

Build Small Hackathon · Backyard AI · June 2026

Badge stack (all merit badges)

Badge	How we earn it
Well-Tuned	LoRA SFT published: GuusBouwensNL/plane-mode-nemotron-4b-study-coach
Llama Champion	Fine-tuned model runs via llama.cpp (./scripts/start_llamacpp.sh + --lora)
Off the Grid	No cloud LLM APIs when PMS_INFERENCE_BACKEND=llamacpp — local llama-server only
Off-Brand	Custom gr.Server UI — plane_mode_scholar/gradio_ui/static/index.html
Sharing is Caring	Open traces: docs/agent-traces-dataset.jsonl + /export_trace
Field Notes	docs/field-notes.md

Full local stack (demo video path)

# 1. Convert HF LoRA → GGUF LoRA (uses llama.cpp convert_lora_to_gguf.py)
python scripts/export_lora_gguf.py

# 2. Start llama-server: Nemotron 4B Q4_K_M + your study-coach LoRA
./scripts/start_llamacpp.sh

# 3. App (second terminal)
PMS_INFERENCE_BACKEND=llamacpp python app.py

Health check should show "inference_backend": "llamacpp" and "lora_applied": true.

HF Space (public demo)

Gradio SDK (default): ZeroGPU + PEFT transformers — Well-Tuned badge on the public Space.

Docker SDK (optional): Embedded llama-cpp-python with GGUF + LoRA — all three local badges on Space. See docs/space-llamacpp.md.

Browser WebGPU: Llamas on the Web proves llama.cpp in the browser; LoRA in wllama is still roadmap — use merged GGUF or screen-record the local llama-server flow for judge demo.

Evidence table

Badge	Status	Evidence
Off the Grid	✅	PMS_INFERENCE_BACKEND=llamacpp → local llama-server, no OpenAI/Anthropic APIs
Off-Brand	✅	Custom gradio.Server frontend inspired by SwarmGrid
Llama Champion	✅	scripts/start_llamacpp.sh, scripts/export_lora_gguf.py, core/llm_llamacpp.py
Sharing is Caring	✅	docs/agent-traces-dataset.jsonl + /export_trace
Field Notes	✅	docs/field-notes.md
Well-Tuned	✅	GuusBouwensNL/plane-mode-nemotron-4b-study-coach
Best Agent (award)	✅	StudyAgent monitor→plan→act — FLY button
Nemotron Quest	✅	Nemotron 3 Nano 4B (fine-tuned) + 30B fallback

Quick verification

curl -s localhost:7860/api/health | python3 -m json.tool
python scripts/export_lora_gguf.py --dry-run
python scripts/verify_finetuned_model.py

UI modes

Mode	Env	Use case
Server (default)	PMS_SERVER_UI=true	Hackathon demo — SwarmGrid-style dashboard
Blocks (legacy)	PMS_USE_BLOCKS=true	Full multi-tab feature surface
llama.cpp	PMS_INFERENCE_BACKEND=llamacpp	Badge stack: Well-Tuned + Llama Champion + Off the Grid