Nam-toon-studio/amrit-qwythos-corrections

Amrit-Qwythos Corrections Dataset

49 question → correct-answer pairs used to fine-tune Qwythos-9B (a Qwen3.5-based model) to fix specific, verified hallucinations caught during real use in the Amrit OS project.

What this fixes

Two real, reproduced failure classes:

1. 1.Self-identity confusion — at higher sampling temperature, the base model would sometimes answer "I am Qwen3.5, made by Alibaba Cloud" instead of its actual fine-tuned identity (Qwythos, by Empero AI).
2. 2.Punjabi/Sikh/Indian-classical-music factual errors — e.g. confusing "tabla" (a musical instrument) with "table" (a data structure), or misexplaining the Sikh greeting "Sat Sri Akal" as a Sanskrit prayer. Includes verified facts about swara/saptak/taal (Hindustani music theory).

Provenance (honest, per-row)

Each row's source field says where it came from:

• —live_qwythos_self_play — a real wrong answer Qwythos gave, caught live by general_play.py (a self-play harness in Amrit OS that asks Qwythos questions with known-correct answers and verifies the reply), paired with the verified correct answer (never the model's own wrong output).
• —authored_correction — supplementary examples for the same failure classes, hand-verified, added because the live-caught set alone (2 rows) was too small to fine-tune on without overfitting.

What this dataset does NOT include

The full Qwythos fine-tune this dataset was part of also used real STEM/medical benchmarks — not re-uploaded here because of their own licenses:

• —GSM8K (MIT) — grade-school math
• —SciQ (CC-BY-NC-3.0) — science Q&A
• —ARC-Challenge (CC-BY-SA-4.0) — science reasoning
• —MedQA-USMLE — US medical licensing exam questions
• —OpenBookQA (Apache-2.0) — science reasoning

Load those directly from their own repos if you want to reproduce the full mix; this dataset is only the part Amrit OS itself generated/verified.

Format

{"question": "...", "answer": "...", "source": "live_qwythos_self_play"}

Honesty note

This is a small, targeted correction set — it teaches a specific model to stop making specific caught mistakes. It is not a general-purpose instruction-tuning dataset and was not designed to be one.