timidan/oya-wallet-agent-data

Oya wallet agent data

The data, logs and benchmarks behind Oya, a LoRA fine-tune of Cohere Labs' Tiny Aya that works as a wallet agent in English, Nigerian Pidgin, Yoruba and Yoruba-English mix. It buys airtime and data, pays for electricity and sends money through seven tools, and it has to ask when details are missing, refuse to overspend, and ignore instructions planted in tool results. Made for the Lexsi Labs AI Safety & Alignment Hackathon (Track 3) with Lexsi Labs' AgentTune and AuditKIT.

Every tool runs against a mock wallet. All names, phone numbers, bank accounts and meter numbers are made up.

Contents

Config or folder	Rows	Contents
sft (train, validation)	2,080 / 104	The reference chats v2 was trained on: compact system prompt, request, tool calls, tool results, final reply
tasks (train, dev, native_speaker, online)	2,080 / 416 / 30 / 18	Each task's starting wallet, expected outcome and reference plan. dev uses phrasings kept out of training; native_speaker holds 30 Yoruba and Pidgin requests typed by a native speaker; online is built from native-written CC BY 4.0 text (INJONGO's Yoruba, Wazobia's Pidgin), each line's source in tasks/online_requests.txt. None were trained on
dpo_pairs (round1, round2)	222 / 146	Pairs from v2's own mistakes: the chat so far, the reference turn (chosen) and a sampled v2 turn that got it wrong (rejected). Round 2 leaves out pairs whose only fault was replying in the wrong language
runs/		Every run in the write-up: trajectories.jsonl in AgentTune's format, per-task scores.jsonl, metrics.json, reports and notebook logs
voice/		The speech models' transcripts of one speaker's recordings of the native-speaker requests (no audio)

Runs by question:

• Untrained vs fine-tuned, 416 development tasks: baseline-compact, sft-v1, and v2 in sft-v2-dev-vs-dpo-v3 (v2 re-run next to the DPO rounds, with the scores the write-up reports).
• The 30 native-speaker tasks: baseline-compact-human, sft-v2-human; the quantized v2 in gguf-*; OpenAI and open models in compare-* and compare.md.
• The online set: baseline-compact-human2, sft-v2-human2, dpo-v3b-human2.
• The two DPO rounds: dpo-v3, dpo-v3b (pairs, config, training stats) and their -dev and -human evaluations.
• The voice test: voice-*.

How it was made

sft and tasks come from the generator in Timidan/lexsi-hack:

OYA_PROMPT=compact python scripts/build_data.py --train-per-cell 40

That is 13 categories × 4 languages × 40 tasks for training, seed 2026. Scores in runs/ are as each run saved them; scripts/rescore.py in the repo applies the current, stricter scorer, which changes only DPO round 1's development score (94.7% to 93.5%). The generator writes the reference turns from templates; no model wrote them. The native-speaker requests as typed are in tasks/native_speaker_requests.txt.

Licence

CC-BY 4.0. Model outputs in runs/ and the rejected turns in dpo_pairs come from Tiny Aya fine-tunes (CC-BY-NC 4.0), OpenAI models and other open models, and stay under their providers' terms. The code is MIT; the Oya models are CC-BY-NC 4.0, following Tiny Aya.

Caveats

• The generator's Yoruba and Pidgin templates are a first draft; native-speaker review is in progress (docs/language_review.md in the repo).
• The 30 native-speaker requests and the voice recordings come from one speaker.