A QLoRA fine-tune of Qwen/Qwen2.5-7B-Instruct for generating Vietnamese Tử Vi Đẩu Số (Chinese/Vietnamese astrology) chart readings, from the PandaScope project (self-hosted Tử Vi web app). Quantized to Q4_K_M (~4.4GB) for easy local use with Ollama:

Trained on 300 template-drafted (not expert-edited) example readings — see the project's training/LEDGER.md for the full training log. It failed its own 0.98 grounding-accuracy gate at 0.9726 on a 12-case benchmark (occasionally states a star/position not actually present in the input chart). Not suitable for serving real astrological decisions — this is shared for the PandaScope app's own local/hobby self-hosting, where the app's own verify.py grounding checker catches and falls back on ungrounded output before showing it to a user. Don't use this checkpoint's output as a standalone source of truth.

It's also only been trained on the one-shot narrative reading format, not on tool-calling — served through plain Ollama, the chat/tool-calling feature can occasionally misfire (rare). PandaScope's own backend/serving/ setup serves this same checkpoint through vLLM instead, which fixes that specific issue (see that project's backend/serving/README.md) — this GGUF build is for the simpler, zero-GPU-required local Ollama path.

Apache 2.0, inherited from the Qwen2.5-7B-Instruct base model.