I'm Not My Parents: Does Improving Parent Language Capabilities Transfer Alignment to Lower Resource Language?

This is an exceptionally strong hackathon project that tackles a highly relevant and under-explored area of AI safety: whether alignment transfers from high-resource "parent" languages to lower-resource regional dialects.

Strengths:

- Your methodology of manually distinguishing between capability failures ("misreads") and true alignment failures ("compliance") is excellent and absolutely critical for low-resource safety evaluations. Too many automated benchmarks conflate the two. Furthermore, adding a mechanistic interpretability layer (logit patching and attention mapping) elevates this project significantly above standard behavioral red-teaming.

Areas for Improvement:

Scale: While $N=103$ is reasonable for a hackathon's time constraints, expanding this to the full dataset with automated (but verified) LLM-as-a-judge pipelines would strengthen the statistical claims.Mechanistic Scope: The behavioral findings highlighted Sahabat-AI's stark drop in performance, but the mechanistic probes were limited to the smaller Qwen3-1.7B. Exporting attention and applying patching to the Llama-3 based Sahabat model would directly tie your strongest behavioral results to your mechanistic theories.Data Visualization: The 3D ENG-shadow attention plot (Figure 4) is a bit difficult to interpret at a glance. Consider sticking to 2D heatmaps (like Figures 2 and 3) or line plots for future publications to improve readability.

Your project asks whether an AI model's safety training carries over from a major language to a smaller related one, testing Indonesian against Acehnese. It's a strong question with a clear product lesson. Safety can look solid in the language a team tests and quietly break in one they don't, which is a real risk for anyone shipping a model across many languages. You went past showing the gap and traced it inside the model, then partly fixed it, which is impressive for the time you had. The repo was private when I reviewed, so I couldn't confirm the deeper claims, and a gated or redacted version would let reviewers check the strongest part of your work. The internal analysis also runs on a small model, so I'd be careful assuming it holds on larger ones until you test that.

Well-scoped work on a real blind spot. The best idea here is separating misreads from true compliance, which changes how you should read a high attack-success rate in a low-resource language. Rating all 412 responses by hand was the right call for this setting. The weak point is that the whole argument rests on translation quality and the misread-vs-compliance label, yet there's no fluent Acehnese check, only machine translation fixed by one author plus LLM-assisted rating. Since translation quality is the experiment, a native review of even a subset is the top priority. The model comparison is also confounded (1.7B vs 8B, full precision vs Q4), so present them as two cases rather than a head-to-head. The mechanistic section is promising but thin: eight exemplars, one striking patching result, Qwen only. Run it across more prompts and apply it to Sahabat, the model you care about. The attention figures are hard to read and barely used; tell the reader what to look for or drop them.