--- title: Hackathon Advisor emoji: "📜" colorFrom: yellow colorTo: green sdk: gradio sdk_version: 6.16.0 python_version: "3.11" app_file: app.py pinned: true license: mit short_description: Originality advisor for small-model project ideas. tags: - gradio - build-small-hackathon - small-models - agent - originality - off-the-grid models: - openbmb/MiniCPM5-1B - build-small-hackathon/hackathon-advisor-minicpm5-lora - build-small-hackathon/hackathon-advisor-quest-minicpm5-lora - ggml-org/embeddinggemma-300m-qat-q8_0-GGUF - nvidia/nemotron-speech-streaming-en-0.6b datasets: - build-small-hackathon/hackathon-advisor-quest-dataset - build-small-hackathon/hackathon-advisor-codex-traces --- # Hackathon Advisor **Hackathon Advisor** is a live map of the Build Small Hackathon and a small-model originality coach for builders. It opens on an atlas of public `build-small-hackathon` Spaces, then lets a builder search the field, inspect project clusters, see quest evidence, and open **The Unwritten Almanac** to evaluate an idea against the work already on the trail. The [Build Small Hackathon](https://huggingface.co/build-small-hackathon) asks participants to build under a 32B parameter cap, solve a concrete problem for someone nearby or make a delightful AI-native experience, and submit a Space, demo video, and social post. Hackathon Advisor treats that setting as the data surface: every public Space becomes part of a continuously refreshed project atlas, and every advisor response is grounded in that shared map. ## Demo - Live app: - Hugging Face Space: - Source code (GitHub): - Demo video: _TODO — add the hosted demo video URL before submission._ - Social post: _TODO — add the public X/LinkedIn post URL before submission._ - Start at the Idea Map, search for a theme, click nearby projects, hover quest badges for evidence, and open the advisor when you are ready to test an idea. ## What This Establishes Builders enter a fast-moving hackathon with limited context. A promising idea can already be crowded, a quiet niche can be hard to see, and prize alignment can be scattered across READMEs, tags, and app files. Hackathon Advisor turns the field itself into the starting point. The app shows where projects cluster, which submissions sit near each other, which quests they appear to satisfy, and where a new idea may still have room to breathe. The atlas is the default experience because the map is the evidence. The advisor is available behind `Open advisor`, where it uses the same project snapshot to cite overlap, propose whitespace, score the idea, draft a build plan, and export the session evidence. ## What You Can Do - Explore a full-screen t-SNE atlas of public hackathon Spaces, with KMeans clusters and nearest-neighbor links. - Search projects with BM25 over titles, slugs, summaries, tags, declared models, cluster labels, quest evidence, README text, and declared app-file source. - Filter by cluster or quest, then inspect the selected project's summary, Space link, tags, quest matches, and evidence hints. - Refresh the atlas from the Space backend; validated artifacts are written to the mounted cache directory and swapped into the live app atomically. - Open the advisor workspace for idea comparison, gap exploration, score seals, profile-aware plans, voice input, and shareable exports. - Export from the workspace UI: build notes, the Almanac chapter, and the page PNG. Further reviewer artifacts — trace JSONL, demo bundle, submission packet, LoRA dataset, and LoRA training kit — are served through the API endpoints listed below. ## How It Works The refresh path snapshots public Spaces in the `build-small-hackathon` organization, reads each README and declared main app file, rebuilds the EmbeddingGemma project index, analyzes quest evidence with MiniCPM, and generates the dashboard payload. The active dashboard contains project points, nearest links, clusters, quest coverage, provenance, and refresh state. `ADVISOR_CACHE_DIR` is the artifact store. On Hugging Face Spaces it points to the mounted Storage Bucket; locally it can be a normal directory such as `.cache/advisor-dashboard`. Each refresh writes `runs/{run_id}/projects.json`, `project_index.json`, `dashboard.json`, `quest_analysis.json`, and `manifest.json`, then updates `latest.json` through an atomic swap. Quest analysis is cached per project using the rendered README+app-file prompt hash, taxonomy hash, MiniCPM model id, adapter id/revision, local adapter digest, and generation config. The app starts an hourly scheduler when `ADVISOR_CACHE_DIR` is configured. Manual and scheduled refreshes both acquire `$ADVISOR_CACHE_DIR/refresh.lock`, heartbeat while active, and leave the current validated dashboard in place if a new run fails validation. ## Models And Data | Role | Model | Runtime | Evidence | | --- | --- | --- | --- | | Advisor | [`openbmb/MiniCPM5-1B`](https://huggingface.co/openbmb/MiniCPM5-1B) + [`build-small-hackathon/hackathon-advisor-minicpm5-lora`](https://huggingface.co/build-small-hackathon/hackathon-advisor-minicpm5-lora) | ZeroGPU, Transformers, PEFT | A 1.08B OpenBMB model plans which tool to call each turn; advisor prose is rendered from deterministic templates grounded in the retrieved tool results. | | Quest analysis | [`openbmb/MiniCPM5-1B`](https://huggingface.co/openbmb/MiniCPM5-1B) + [`build-small-hackathon/hackathon-advisor-quest-minicpm5-lora`](https://huggingface.co/build-small-hackathon/hackathon-advisor-quest-minicpm5-lora) | ZeroGPU, Transformers, PEFT | A task-specific MiniCPM LoRA classifies README and app-file evidence into strict quest JSON. | | Project retrieval | [`ggml-org/embeddinggemma-300m-qat-q8_0-GGUF`](https://huggingface.co/ggml-org/embeddinggemma-300m-qat-q8_0-GGUF) | Local llama.cpp index build plus llama.cpp query embeddings | The atlas and retrieval index use a GGUF embedding model through llama.cpp. | | Voice input | [`nvidia/nemotron-speech-streaming-en-0.6b`](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b) | ZeroGPU; NVIDIA NeMo ASR | Voice notes are transcribed with NVIDIA NeMo using the same Nemotron model in local and deployed runs. | MiniCPM is loaded following the official demo shape (`trust_remote_code=True`, `bfloat16`, and `apply_chat_template(..., enable_thinking=False)`) for stable tool calls and strict quest JSON. | Data / released material | Link | How it is used | | --- | --- | --- | | Hackathon project corpus | [`build-small-hackathon`](https://huggingface.co/build-small-hackathon) | Public Spaces are crawled as the live field for the atlas, search, advisor citations, and quest coverage. | | Project snapshot | [`data/projects.json`](https://huggingface.co/spaces/build-small-hackathon/hackathon-advisor/blob/main/data/projects.json) | Stores Space metadata, README text, declared models/datasets, tags, and declared app-file evidence. | | Project embedding index | [`data/project_index.json`](https://huggingface.co/spaces/build-small-hackathon/hackathon-advisor/blob/main/data/project_index.json) | Stores normalized EmbeddingGemma vectors and retrieval metadata for map construction and advisor search. | | Quest SFT dataset | [`build-small-hackathon/hackathon-advisor-quest-dataset`](https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-quest-dataset) | Trains the MiniCPM quest classifier from README/app-file prompts with source-attributed quest labels. | | Codex session traces | [`build-small-hackathon/hackathon-advisor-codex-traces`](https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-codex-traces) | Publishes real Codex session logs for this project after selection, minimization, and OpenAI Privacy Filter redaction. | | Advisor LoRA examples | `lora_dataset` and [`/api/lora-training-kit.zip`](https://build-small-hackathon-hackathon-advisor.hf.space/api/lora-training-kit.zip) | Regenerates chat JSONL examples, recipe metadata, and the adapter card from exact advisor sessions. | ## How Codex Was Used [Codex](https://developers.openai.com/codex) served as the engineering partner for the project. It helped translate the hackathon requirements into implementation slices, inspect the existing codebase, build the atlas refresh/storage/cache path, add the dashboard search and quest-evidence UI, run local tests and browser checks, review deployed Space behavior, prepare commits and deployment updates, and revise the README into a submission narrative. The live app runtime uses the models and data listed above; Codex appears in the development record as the assistant that helped design, implement, validate, and document the system. The redacted session-level Codex traces are published as a Hugging Face dataset at [`build-small-hackathon/hackathon-advisor-codex-traces`](https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-codex-traces). The full development history is public at . ## Prize Evidence This submission targets the **Thousand Token Wood** main track, plus the OpenBMB, OpenAI/Codex, NVIDIA, and Modal sponsor awards and the six bonus-quest badges. | Prize path | Implemented evidence | | --- | --- | | Thousand Token Wood | The Almanac and Idea Map make the AI output visible as a playful, evidence-grounded exploration surface; the embedding index and the MiniCPM tool loop are load-bearing for the whitespace and originality experience. | | Off the Grid | Every model runs from open weights on the Space's own GPU/CPU (or a local box); no third-party inference API is called at runtime, and retrieval vectors are local and embedded through llama.cpp. | | Well-Tuned | Two MiniCPM5-1B PEFT LoRA adapters (advisor + quest classifier) are published publicly on the Hub; the local quest adapter is byte-identical to its published repo, and the training kit reproduces them. | | Off-Brand | The custom `gradio.Server` frontend ships a bespoke atlas and Almanac experience, with no default Gradio UI in the runtime path. | | Llama Champion | EmbeddingGemma GGUF vectors and every runtime query embedding run through llama.cpp; the index validator rejects any non-llama.cpp runtime. | | Sharing is Caring | Real Codex session logs for this project are published on the Hub at [`build-small-hackathon/hackathon-advisor-codex-traces`](https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-codex-traces); the publisher selects project-relevant sessions, minimizes internal metadata, applies [`openai/privacy-filter`](https://huggingface.co/openai/privacy-filter), and records source hashes for audit. | | Field Notes | A build report on the quest-classifier fine-tune is published at [`docs/quest-classification-lora.md`](docs/quest-classification-lora.md), and the app exports session Field Notes as markdown. | | Tiny Titan | The largest single model is MiniCPM5-1B at ~1.08B — well under the 4B Tiny Titan ceiling; the full runtime stack totals ≈1.98B, far under the 32B cap. | | OpenBMB | MiniCPM5-1B is the central language model for both tool planning and quest classification. | | NVIDIA Nemotron | Voice input runs `nvidia/nemotron-speech-streaming-en-0.6b` through NVIDIA NeMo. | | Modal | Modal trains the quest-classifier LoRA (`scripts/modal_train_quest_lora.py`), and a Modal remote index-build path is provided; the index shipped in this repo was built locally. | | Best Agent | Each turn MiniCPM5 selects one tool; the engine then orchestrates the search → whitespace → score → plan chain over the live project field. | ## Run Locally ```bash python3.11 -m venv .venv . .venv/bin/activate pip install -r requirements.txt mkdir -p .cache/advisor-dashboard ADVISOR_CACHE_DIR=.cache/advisor-dashboard \ ADVISOR_MODEL_BACKEND=minicpm-transformers \ ADVISOR_MODEL_ID=openbmb/MiniCPM5-1B \ ADVISOR_ADAPTER_ID=build-small-hackathon/hackathon-advisor-minicpm5-lora \ ADVISOR_ADAPTER_REVISION=25de69bcde397e1bcdd852923b56a42f10222650 \ ADVISOR_QUEST_ANALYZER_BACKEND=minicpm-transformers \ python app.py ``` Then open . The atlas refresh button runs locally with the same artifact swap path used in deployment. It writes refreshed runs under `.cache/advisor-dashboard/runs/` and atomically updates `.cache/advisor-dashboard/latest.json`. ## Refresh The Project Snapshot ```bash python scripts/crawl_hf_spaces.py --org build-small-hackathon --out data/projects.json python scripts/build_project_index.py --location modal --projects data/projects.json --out data/project_index.json ``` The checked-in development snapshot lives in `data/projects.json` and `data/project_index.json`. A configured `ADVISOR_CACHE_DIR` supplies the latest validated dashboard artifacts. ## Publish Codex Trace Dataset Local privacy-filter run: ```bash uv run --with 'transformers>=5.6,<6' --with 'torch>=2.8,<3' \ python scripts/publish_codex_trace_dataset.py \ --project-root . \ --repo-id build-small-hackathon/hackathon-advisor-codex-traces \ --verbose ``` Faster Modal GPU run: ```bash python scripts/publish_codex_trace_dataset.py --location modal \ --project-root . \ --repo-id build-small-hackathon/hackathon-advisor-codex-traces ``` The publisher scans `~/.codex/sessions` and `~/.codex/archived_sessions`, selects sessions that mention this project, keeps project-facing Codex events, removes system/developer prompts and compaction internals, normalizes local paths, caps long tool-output text with truncation counts in the manifest, applies OpenAI Privacy Filter to the published log text, writes `codex_sessions.jsonl` and `dataset_manifest.json`, then uploads the filtered data to the configured Hugging Face dataset. The Modal wrapper uploads the selected raw JSONL files to a private Modal Volume, runs the same publisher core on a GPU, returns the filtered dataset to local disk, and performs the Hugging Face upload from local credentials. ## API And Artifacts | Surface | Purpose | | --- | --- | | `GET /api/dashboard` | Atlas points, links, clusters, quest report, provenance, and refresh status. | | `GET /api/dashboard/search?q=...` | BM25 search over project, cluster, quest, README, and app-file text. | | `POST /api/dashboard/refresh` | Starts one background refresh job. | | `GET /api/dashboard/refresh` | Reports refresh stage, result, and status. | | `POST /api/transcribe` | Transcribes uploaded voice notes with NVIDIA NeMo and Nemotron ASR. | | `GET /api/prize-ledger` | Model stack, parameter budget, runtime status, and prize evidence. | | `GET /api/demo-bundle.zip` | Demo session JSON, prize ledger, trace, notes, chapter, LoRA files, submission packet, and PNG. | | `GET /api/lora-training-kit.zip` | SFT data, recipe, adapter card, and training command. | The Gradio API also exposes `trace_artifact`, `field_notes`, `chapter`, `lora_dataset`, and `submission_packet` for submission evidence and reviewer inspection. ## Advisor Workspace The advisor workspace preserves the working loop from the original app. `Ink` compares the current idea against the project index, `Gap` rotates through unused whitespace candidates, `Plan` drafts a practical build path, and `Compare` rescans the saved idea board to select the strongest page. The `Profile` panel adds skills, time, preferences, and constraints to the plan so the output can reflect "one evening", "frontend prototyping", or "CPU-only Space" as real scoping facts. Each scored page includes a deterministic `wood_map`: background dots for indexed Spaces, red dots for closest cited echoes, and a green/red point for the current idea. The live UI and PNG export use the same Pillow renderer. ## Runtime Backend The deployed Space is configured for ZeroGPU inference with: ```bash ADVISOR_ZERO_GPU=1 ADVISOR_ZERO_GPU_DURATION=120 ADVISOR_MODEL_BACKEND=minicpm-transformers ADVISOR_MODEL_ID=openbmb/MiniCPM5-1B ADVISOR_ADAPTER_ID=build-small-hackathon/hackathon-advisor-minicpm5-lora ADVISOR_ADAPTER_REVISION=25de69bcde397e1bcdd852923b56a42f10222650 ADVISOR_QUEST_ANALYZER_BACKEND=minicpm-transformers ADVISOR_QUEST_ADAPTER_ID=build-small-hackathon/hackathon-advisor-quest-minicpm5-lora ADVISOR_QUEST_ANALYSIS_BATCH_SIZE=8 ADVISOR_CACHE_DIR=/data/advisor-cache ADVISOR_REFRESH_COMPUTE=cpu ADVISOR_SCHEDULED_REFRESH_COMPUTE=cpu ADVISOR_REFRESH_INTERVAL_SECONDS=3600 ADVISOR_REFRESH_INITIAL_DELAY_SECONDS=300 ADVISOR_REFRESH_LOCK_TTL_SECONDS=7200 ADVISOR_REFRESH_EMBEDDING_TIMEOUT_SECONDS=1800 ADVISOR_EMBEDDING_MODEL_REPO=ggml-org/embeddinggemma-300m-qat-q8_0-GGUF ADVISOR_EMBEDDING_MODEL_FILE=embeddinggemma-300m-qat-Q8_0.gguf ADVISOR_EMBEDDING_N_CTX=2048 ADVISOR_ASR_MODEL_ID=nvidia/nemotron-speech-streaming-en-0.6b ``` The retrieval query embedder downloads the GGUF model through `huggingface_hub` unless `ADVISOR_EMBEDDING_MODEL_PATH` points to a local file. `/api/transcribe` uses the same ZeroGPU wrapper for Nemotron ASR. On macOS local runs, the app automatically runs llama.cpp query embedding in a worker process so the MiniCPM PyTorch runtime and llama.cpp stay isolated from each other's OpenMP runtime. Dashboard refresh also builds the GGUF embedding index in a subprocess before returning to the app process for MiniCPM quest analysis. When `ADVISOR_CACHE_DIR` is set and `HF_HOME` is not, the refresh subprocess stores Hugging Face downloads under `$ADVISOR_CACHE_DIR/huggingface` so the mounted bucket keeps the embedding model cache across refreshes and restarts. ## Test ```bash pytest ```