Muslim-6B-v3 (test/candidate build — pending eval-gate confirmation)

This is a candidate third iteration of NightPrince/Muslim-6B (v1) and NightPrince/Muslim-6B-v2, deployed for live comparison testing. The Phase 6 eval gate (regression + held-out generalization + repetition-loop checks) may still be running when this was pushed — check the project's eval logs before treating this as a confirmed release.

What changed from v2

• Fixed three verified tool-schema bugs, found by live-querying the real mcp.tafsir.net MCP server and inspecting real captured production tool calls (not guessed): analyze_word restructured from a {word} param (which doesn't exist on the real tool) to the real {surah, ayah, word_no} shape; fetch_nuzool_reason now includes the real tool's required ayah param; fetch_hadith's param renamed collection→collection_slug to match the real tool and the already-correct search_hadith.
• Added live-tested generalization fixes: surah nickname resolution (قلب القرآن, عروس القرآن, etc.), English surah-name resolution, explicit reciter-namespace disambiguation (a single request needing both a play_ayah-format and a play_surah-format reciter key, via new multi-tool training traces), broader B5 ruling coverage (sihr/kahana), and more diverse English creator-identity phrasings.
• New B7 behavior: graceful decline when a tool call errors or no tool result is available, instead of confidently confabulating — added after live testing found v1 and v2 both guessed wrong surah/ayah numbers more confidently than the untrained base when no tool was bound.
• Training loss decreased monotonically across all 3 epochs (0.157 → 0.144 → 0.140), unlike v2 which showed a mild epoch-3 uptick — no overfitting signal in the loss curve. (Per this project's own experience, a clean loss curve is necessary but not sufficient — see the eval gate results below.)

Status

Built and trained following a deliberate re-grounding process: every fix here traces to a specific, live-confirmed failure (real MCP schema verification, a 36-probe live comparison across base/v1/v2), not a guess. A held-out probe set (eval/probe_prompts_v3.py) — covering nicknames, English names, and scenarios genuinely absent from training — was built specifically to test generalization before declaring success. Check the repo's logs/eval_gate_v3.log for the actual pass/fail result before relying on this model.

License

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