Muslim-6B-v2 (test/candidate build — not yet the recommended release)

This is a candidate second iteration of NightPrince/Muslim-6B, deployed separately for live comparison testing against the v1 release and the untuned Karnak-6B-v1.0 base. It has not completed the full evaluation process and should not be treated as a final replacement for v1.

What changed from v1

• Training data expanded from 316 to 393 examples (362 train / 31 val), specifically adding more surah-name/ayah-number coverage (B1) and more measured-ruling examples (B5) to address two caveats documented in the v1 model card.
• Same QLoRA recipe: r=16, alpha=32, dropout=0.05, targeting q/k/v/o/gate/up/down_proj, 3 epochs, fp16 compute on an RTX 2080 Ti.

What we've verified so far

On the same 18-probe set used for v1, the two v1 caveats (Surah Yusuf mis-mapped to surah 34; unconditional takfir framing on "one who abandons prayer") were fixed, and the fix generalized correctly to held-out, never-trained surahs and rephrasings in a follow-up 12-probe generalization set.

That same generalization set also surfaced two new, previously-undetected issues, which is why this is a candidate build, not a release:

• Reciter-argument confusion. play_ayah and play_surah use two different reciter-key vocabularies in the real production tool schemas; this model's output sometimes uses the wrong one for the tool being called (e.g. an ayah-tool reciter key passed to the surah tool).
• Nickname / English surah-name resolution is weak. A well-known Arabic nickname for a surah ("قلب القرآن") and an English surah name both resolved to the wrong surah number, even though the same skill worked correctly for Arabic canonical names.
• Ruling answers can still volunteer hudud-adjacent language unprompted on ruling topics outside the training set (tested: sorcery/sihr) — better than the untuned base, but not a clean pass.

Status

This build exists for side-by-side comparison testing (base Karnak vs. Muslim-6B-v1 vs. Muslim-6B-v2) as part of a broader re-evaluation against the real production system prompt and real MCP tool servers. Do not deploy this as a production brain ahead of that evaluation completing.

License

Apache 2.0, inherited from the base model lineage (Qwen3-4B-Instruct-2507 → Karnak-6B-v1.0).