# LaneMint

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aaron Alsop](https://cerebralvalley.ai/u/AGV)
- **GitHub:** https://github.com/alsopgroup/lanemint-meta-hackathon
- **Demo video:** https://youtu.be/vZs__XZ4xKo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/173

LaneMint builds hard, verifiable quant-finance tasks for training and evaluating AI models. Real market data is messy: bad price spikes, duplicate rows, missing days, unadjusted stock splits. A model that misses one of these gives a confident answer that's wrong, and in finance that costs real money.

For this hackathon we built two things on Meta's models. The whole project was built with Meta's models. Every coding agent that wrote this repo ran on Muse Wonder, Muse Spark powers the live app, and Muse Glimmer and Spark Open are the open models we measured and fine-tuned.

1. A live data-audit app at https://meta.lanemint.com. You upload a price file, and Muse Spark (through the Meta Model API) finds the problems in it, explains each one and shows a cleaned "after" view. For the public sample tasks, a deterministic scorer then grades the result.

2. An open benchmark, tools_frontier. These are multi-turn, tool-using data tasks generated fresh from a seed, so the answers can't be memorised. We use it to find where Meta's models stop being reliable. On held-out seeds with a 12-turn budget, Muse Spark 1.3 solved 70%, Muse Spark Open 57% and Muse Glimmer 30%. Most of Spark's misses came from running out of turns, not from wrong reasoning. We also tried LoRA fine-tuning on Glimmer through Fireworks. Within the hackathon's time and compute window, those short training runs didn't improve the base model. We're publishing that as an honest result and a starting point for longer training after the event.

Try it in two minutes: the repo's samples/ folder has three ready-made tasks with answer keys, which you can upload to the live app.

Everything in this repo was built during the hackathon. It's a new repository created on kickoff day (Oct 26), and it contains the app, the tools_frontier benchmark and its task generator, the Muse model evals, the Glimmer LoRA experiments and the public samples. The design draws on lessons from LaneMint's earlier work, but no code, tasks or data from it are included.

Run it yourself with Docker: clone the repo and run `docker compose up --build app`, then open http://localhost:8000. Demo mode works with no API keys, and a sample benchmark run is included. See "Run with Docker" in the README.

## More from Meta Global AI Developer Hackathon

- [Shamuel Perez](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/170)
- [Hoist - Decision Canvas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/171)
- [DriveIQ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/172)
- [Luscher](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/174)
- [flowcheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/175)
- [lightgo1230](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/176)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/173. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt