Ìdìbò: Tiny Aya Earth for safe voter information (Nigeria 2027)

A LoRA fine-tune of CohereLabs/tiny-aya-earth (3.35B) that answers voters' questions in English, Nigerian Pidgin, Yorùbá and Hausa (Igbo: beta, untrained). It answers from retrieved INEC and Electoral Act 2026 passages, with citations. It refuses election harm, stays politically neutral, abstains when it doesn't know, and still helps with legitimate civic questions.

Demo video: https://youtu.be/Nn0A--cZJWw · Try it: https://idibo-2027.up.railway.app · Code: https://github.com/Gbangbolaoluwagbemiga/idibo · Write-up: https://github.com/Gbangbolaoluwagbemiga/idibo/blob/main/docs/WRITEUP.md · Results page: https://idibo-2027.up.railway.app/results.html

Files

File	What
`adapters/idibo/`	Ìdìbò: LoRA r16/α32 on all projections; 1,060 RAFT-style examples (civic Q&A, refusals with a lawful alternative, neutral answers, abstentions) in en/pcm/yo
`adapters/civic/`	Control: the same recipe on civic Q&A only (566 examples), the "plain fine-tune"
`adapters/civic-safelora/`	The civic adapter repaired with SafeTune Safe LoRA: no training; 74 of 252 LoRA modules projected
`idibo-q4_k_m.gguf`	Ìdìbò merged and quantised (Q4_K_M, 2.1 GB) for llama.cpp: about 10–14 s per answer on 8 vCPUs

Training: AlignTune RaftTrainer, 2 epochs, lr 1e-4 (cosine), fp16, completion-only loss, each prompt holding the golden passage plus 3 retrieved distractors. Data curated with CuratorKIT. Trained on Kaggle (one T4, 3.1 h).

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-earth")
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-earth"),
                                  "GOPAE/idibo-tiny-aya-earth", subfolder="adapters/idibo")

Use it with the system prompt and retrieval from the repository (scripts/run_eval.py, idibo/retrieve.py). It was trained to answer from SOURCES: passages.

Results (summary; full tables with 95% intervals are in the write-up)

Held-out benchmark v2, English (56 harmful incl. 16 jailbreaks, 48 benign look-alikes), % [95% CI], graded by command-r7b. AuditKIT paired-bootstrap p-values against the plain fine-tune.

	Harmful compliance	Jailbreak compliance	Over-refusal
Base + same prompt and sources	28.6 [18.4–41.5]	50.0 [28.0–72.0]	2.1 [0.4–10.9]
Plain civic fine-tune (`civic`)	0 [0–6.4]	0 [0–19.4]	35.4 [23.4–49.6]
+ SafeTune ReSta, no training	0 [0–6.4]	0 [0–19.4]	27.1 [16.6–41.0] (p = 0.27)
+ SafeTune Safe LoRA, no training (`civic-safelora`)	5.4 [1.8–14.6]	0 [0–19.4]	16.7 [8.7–29.6] (p = 0.012)
Ìdìbò (`idibo`)	0 [0–6.4]	0 [0–19.4]	6.2 [2.1–16.8] (p = 0.001)

Hausa (never trained): harmful 5.6%, over-refusal 22.2% (plain fine-tune 55.6%). Igbo (never trained) is the weakest: harmful 16.7%. The command-r7b grader is lenient on harmful items (κ = 0.28 against command-a-plus), so harmful rates are lower bounds.

Round 1, Pidgin, graded by command-a-plus:

	Base + sources	Plain fine-tune	Ìdìbò
Harmful compliance	55.5%	5.6%	0%
Over-refusal	11.1%	61.1%	16.7%
Took a political side	24.4%	6.7%	0%
Answers in Pidgin	12.8%	100%	100%

Limitations

Yorùbá and Igbo factual accuracy is limited by cross-lingual retrieval. The app translates the question to English before searching, which raised recall@4 from 7→12/20 (yo) and 4→16/20 (ha).
Translations in training and evaluation are machine-made; native-speaker review was partial.
In some refusals the model invents INEC contact details (e.g. a hotline 0800-INEC-HELP, the non-existent domain inec.gov.ng): 6.7% of answers. The deployed app removes any contact not found in the sources; if you use the adapter elsewhere, do the same.
Over-refusal stays high on code-mixed prompts (4 of 6 benign) and in Yorùbá (44%).
The benchmarks are small (v2: 56 harmful, 48 benign, 30 code-mixed). Read the intervals, not just the point estimates.
Not official INEC guidance. Election details change: always check inecnigeria.org.
Non-commercial licence, inherited from Tiny Aya.