Field Atlas is a free, open-source Android app that answers questions offline from sources saved on the phone: an encyclopedia, travel guides and first-aid guides. It started as a poidh bounty. This one is paid from the app's token fees.
The problem
On a 6 GB phone the app runs Qwen3.5 2B or 4B through llama.cpp. Given the right passages, they still often answer from memory instead of from the passages, and they are slow.
The task
Fine-tune a small model that reads the passages it is given, answers from them, cites them, and says plainly when they do not contain the answer.
Requirements
- Base model of 4B parameters or fewer, under an open license (Apache-2.0 or MIT).
- Runs as a GGUF (Q4_K_M or smaller) with llama.cpp on a 6 GB Android phone.
- Input: a question and up to 8 numbered passages. Output: a short answer that cites the passages it used, like
[S2]. If the passages do not answer the question, it says so.
- Never adds first-aid steps that are not in the passages.
Submit (reply here with a link)
- The GGUF on Hugging Face.
- The training script and a description of the training data. Use only data you are allowed to share.
- Your score on the public dev set, from
tools/evaluation/reader_bench.py.
- Everything under an open license.
Judging
- The dev set (60 questions, the passages the app's own search found for each, reference answers) and the scoring script: github.com/0x94t3z/fieldatlas/tree/main/docs/fieldtest/reader
- Every submission is run on a sealed test set of 68 questions built the same way, not published until judging is done. SHA-256:
cf12c16d5bc6ae248130721a77788e524c1d5a36d4d7a5b07285196232da2dce
- Scored on: correct answers supported by the cited passage; saying "not in the sources" when that is true; no unsafe first-aid advice; speed on a 6 GB Android phone.
- To win, a submission must beat the app's current Qwen3.5 2B on the sealed set (it scores 36.7% on the dev set). The best score wins. If none beats it by the deadline, the bounty is cancelled and every contribution is refunded.
- Scores for every submission are published in the repo.
Deadline
2 December 2026