Language models for African languages now have usable base representations, but they still cannot follow instructions or hold a useful conversation. The reason is that almost no instruction data exists in these languages. We propose to collect that data in language classrooms. University students studying isiZulu or isiXhosa work in small facilitator-led groups on locally grounded problems, alongside a language model that acts as a peer. The model sometimes phrases its suggestions tentatively and sometimes makes deliberate small mistakes, which the students must catch and fix. Each logged session gives two kinds of training data: multi-turn dialogues written in the target language, and paired examples in which a wrong model turn sits next to a native speaker's correction. Pairs like these, a rejected and a preferred response, are what preference tuning needs.
The bet is specific. We do not claim that classroom data is simply "more" data. We claim it contains something that paid instruction writing and machine translation rarely produce: conversations in which ambiguity is resolved, errors are noticed, and the lead passes back and forth between speakers. A model tuned on this should be better at the multi-turn, error-recovering behavior that real users need. We test this claim directly against paying native speakers to write instructions at the same budget. That comparison is the one most likely to sink the idea.
Core Claim
Continued pre-training (further training an existing model on raw target-language text) has worked for African languages. AfriqueLLM adapted Qwen 3, Llama 3.1 and Gemma 3 on a 26-billion-token mixture of African text, code, math and synthetic translation. AfriqueQwen-14B reached an evaluation score of 63.58, above Gemma 3 27B at 60.71. The authors state that the remaining bottleneck is the scarcity of localized instruction data. This paper motivates our proposal: it supplies the base model and names the gap, but says nothing about how to fill it.
Our claim is that facilitator-led problem solving with a deliberately fallible model peer yields instruction data that is better per token on multi-turn and error-recovery behavior than equal-budget direct instruction writing. It should also be no worse on standard benchmarks. A second claim is that this data is cheap per usable example, because the learning activity is the incentive. We also expect the gain to be larger on a language-adapted base model than on a generic multilingual one. The reason is that instruction data can only surface abilities the base representations already have.
Where the Pieces Come From
The mechanism for eliciting data comes from KSC-PaL, a peer-learning agent for English computer-science problems. It tracks "initiative shifts," the moments when the lead in a dialogue passes between student and agent. To draw students out, it uses hedging, requests for feedback, and deliberate errors. In its corpus, initiative shifts correlated with learning gains ($R^2 = 0.20$) and occurred more often during joint knowledge-building episodes ($t(57) = 3.32$). KSC-PaL used these tactics only to teach, and it needed a human "wizard" to interpret what students typed. We reuse the tactics for a different purpose: as a device for producing correction data. Our model peer never has to fully understand students, because the facilitator and the students themselves validate what gets logged.
The classroom design follows the problem-based learning study of Exploring Engagement and Efficacy in Secondary English Education in China. That study used facilitator-guided problem generation, groups of 5–7 learners, a mix of in-person and chat interaction, and peer feedback. It reported gains in engagement and self-efficacy, but in a single uncontrolled cohort of 45 students, so we treat it as a template for session design rather than as evidence about outcomes. Neither pedagogy paper connects its work to building training corpora.
The closest data-collection precedent is EnetCollect in Italy. It gathered linguistic resources as a byproduct of language-learning exercises and weighted contributions by learner proficiency. Its outputs were static micro-annotations, such as part-of-speech tags, multi-word expressions and post-edited translations. We differ in two ways: we collect multi-turn, scaffolded dialogue, and we aim the output at instruction and preference tuning of modern language models.
Community annotation efforts such as MasakhaNER and MasakhaNER 2.0 show that native-speaker communities can produce high-agreement data. They label existing news text, however, and do not produce dialogue.
Two further papers constrain the design. The first is Does Corpus Quality Really Matter for Low-Resource Languages?. It found that curated Basque text (66.5 average F1 across five understanding tasks) did no better than noisy crawls (CC100 at 67.9). The curated corpus also lagged on informal social-media tasks, because curation had kept mostly formal registers. The authors suggest that quality should matter more for generation. That paper is our warning against assuming that "curated" means "better." Classroom dialogue is useful partly because it is informal yet clean, and we measure register diversity rather than assuming it. The second is AfriqueLLM's finding that raw monolingual text alone erodes reasoning. For that reason we mix a replay buffer of existing general instruction data into every tuning condition.
Research Program
We focus on isiZulu and isiXhosa, two closely related Nguni languages. Both have a benchmark suite in NGLUEni, which adapted XLM-R and ByT5 on a 260.95M-token Nguni corpus. Both also appear in MasakhaNER 2.0, where transfer from isiXhosa to isiZulu reaches 85.5 F1. Keeping to two related languages lets us check cross-language transfer of the dialogue data cheaply.
Participants are first-year university students enrolled in isiZulu or isiXhosa courses, both home-language speakers and additional-language learners. Using university students avoids the consent issues of working with minors.
Each 90-minute session presents a group with a local problem posed in the target language. Examples include planning a community budget, interpreting a municipal notice, or solving a word problem set in a local market. The model peer starts as AfriqueQwen-8B with light tuning on a few hundred seed dialogues. It contributes steps and follows a fixed policy for hedging, feedback requests and planted errors. Planted errors come in three kinds: a factual slip, an arithmetic slip, and a grammatical or register error typical of machine-generated isiZulu. The facilitator steers groups toward problems at the edge of their ability and resolves disputes. At the end of each session, a home-language speaker reviews and tags the log.
Logs are converted into two products. The first is multi-turn tuning dialogues. The second is preference pairs, each consisting of a planted-error turn (rejected) and the corrected turn (preferred), plus pairs from the model's own unplanted mistakes that students caught. Contributor proficiency is recorded and used as a reliability weight, following EnetCollect. We target roughly 3,000 validated dialogues across the two languages over one academic term. We will also record facilitator hours and cost per retained example, so the cost claim is measured rather than asserted.
Baselines
The naive reference point is AfriqueQwen-8B with no target-language instruction tuning. It is prompted few-shot and receives only the general replay buffer that every other condition also gets. This shows how much any instruction data helps at all.
The strongest competitor is paid direct writing. Home-language isiZulu and isiXhosa speakers are paid to write instruction–response pairs and short multi-turn dialogues on the same problem themes, with the same review. This is the obvious professional way to build instruction data. It isolates whether the classroom adds anything beyond "native speakers wrote it." We run it twice. In the equal-token version, it is matched to the classroom set in token count. In the equal-cost version, it gets the same total human spend, which will likely buy more tokens than the classroom yields. Tuning recipe, replay buffer, hyperparameter search budget and number of runs are identical across conditions.
The third baseline is translate-train: an English instruction corpus machine-translated into both languages and matched in tokens. AfriqueLLM found synthetic translation helpful as replay data but not a substitute for native text. We check whether that holds for instruction tuning as well.
Two ablations isolate the key ingredient. In the first, facilitated sessions run with a model peer that never hedges or plants errors, which yields dialogue without the correction structure and without preference pairs from planted errors. In the second, learners in the same courses write question–answer pairs on their own, without facilitation or a model peer, as in EnetCollect-style explicit crowdsourcing. If full classroom data beats the no-error ablation, the planted-error mechanism is doing work. If it beats unfacilitated writing, facilitation is doing work.
To cross the tuning factor with base-model type, we repeat the main conditions on generic Qwen 3 8B.
On reported numbers, the relevant figure in the file is AfriqueQwen-14B's evaluation score of 63.58 from AfriqueLLM, which we take as the benchmark-level ceiling to preserve. We found no reported per-subtask AfroBench-Lite breakdown on Lacuna, and no reported result on Lacuna for instruction-tuned models on isiZulu or isiXhosa dialogue. Beating the paid-writing baseline means a statistically reliable win on our multi-turn and error-recovery evaluations at equal tokens, while staying within noise of it on AfroBench-Lite and NGLUEni generation.
Why not just…?
Why not just pay native speakers to write instructions? This is the real threat, and the proposal is built around it. It is plausible that careful paid writers produce data as good as classroom logs. Nothing in the file shows otherwise, because no one has compared the two. The equal-token comparison settles the quality question. If paid writing matches classroom data on error recovery, our quality claim fails. What remains is then only the cost claim, which the equal-cost comparison and the per-example cost accounting would test on their own terms. We will report that outcome plainly rather than reframe it.
Why not just translate English instruction data? Translation is cheap, but the file's evidence cuts against it for this use. AfriqueLLM found synthetic translations useful only as a buffer alongside native text. The AfriSenti work (benchmark, shared task) documents code-switching and dialectal variation that machine translation handles poorly. Translation also cannot produce preference pairs grounded in the errors a model actually makes in isiZulu. The translate-train baseline measures the gap directly.
Why not just prompt a frontier multilingual model, or distill from one? For related scripts the evidence is discouraging. SeaLLMs reports ChatGPT-3.5 at 35.8 on translated GSM8K in Thai, against 62.0 for a 7B regional model. It also documents tokenizers that need up to 15 times as many tokens as English for some scripts. We found no reported result on Lacuna for frontier models on isiZulu or isiXhosa instruction following. Distilled data would also inherit exactly the unnatural register that our planted-error category targets. We include few-shot prompting of the strongest accessible frontier model as an unmatched reference point on the evaluation set, not as a data source.
Proposed Experiments
The primary evaluation is a held-out set of multi-turn tasks written by home-language speakers who did not take part in collection. Some tasks contain an ambiguous request. Others contain a user mistake or a mistake the model makes earlier in the conversation. Separate native-speaker judges compare outputs pairwise without knowing which condition produced them. From this we report win rates with bootstrap confidence intervals. We also report an error-recovery rate: the fraction of planted mistakes the model flags or corrects, and the fraction of ambiguous requests where it asks a clarifying question instead of guessing.
Secondary evaluations check that nothing regresses. These are AfroBench-Lite for reasoning, question answering and translation; NGLUEni data-to-text and headline generation, reported in BLEU; and MasakhaNER 2.0 isiZulu and isiXhosa entity recognition. Following the Basque finding about register, we also report a lexical and register-diversity profile of each tuning set.
Three checks decide the outcome. The first is whether full classroom data beats paid writing at equal tokens on error recovery by a margin whose confidence interval excludes zero. The second is whether that advantage disappears in the no-planted-error ablation, which would locate the effect in the correction structure rather than in the classroom setting as such. The third is whether the gain is larger on AfriqueQwen-8B than on generic Qwen 3 8B, which would show that instruction data and language-specific adaptation complement each other. Finally, we tune on isiZulu dialogues only and test on isiXhosa. This shows whether classroom data collected in one Nguni language transfers to a close relative, as entity recognition does, and hence whether a single classroom program could serve the less-resourced isiNdebele and Siswati.
References
- AfriqueLLM: How Data Mixing and Model Architecture Impact Continued Pre-training for African Languages
- KSC-PaL: A Peer Learning Agent that Encourages Students to take the Initiative
- Exploring Engagement and Efficacy in Secondary English Education in China: A Problem-Based Social Constructivism Approach
- EnetCollect in Italy
- MasakhaNER: Named Entity Recognition for African Languages
- MasakhaNER 2.0: Africa-centric Transfer Learning for Named Entity Recognition
- Does Corpus Quality Really Matter for Low-Resource Languages?
- NGLUEni: Benchmarking and Adapting Pretrained Language Models for Nguni Languages
- AfriSenti: A Twitter Sentiment Analysis Benchmark for African Languages
- SemEval-2023 Task 12: Sentiment Analysis for African Languages (AfriSenti-SemEval)
- SeaLLMs -- Large Language Models for Southeast Asia