A 1B model fine-tuned to read Chinese BaZi (八字) / I-Ching (六爻) charts in a warm, plain, second-person, anti-doom voice — translating esoteric symbols into everyday psychological language and always ending with one concrete next step. It powers the Tianwen app for the Build Small Hackathon.

Base model: openbmb/MiniCPM5-1B
Method: LoRA (rank 16, alpha 32, dropout 0.05, target=all), bf16, template="empty", cutoff_len=4096
Data: 58 distilled samples (teacher: MiniMax-M2.7-highspeed) — see tianwen-distill
Training: Modal A100, 8 epochs, lr 2e-4 cosine — train_loss 1.939, loss 3.5 → 1.0 in 91s
Formats: F16 GGUF (tianwen-minicpm1b.gguf, 2.1 GB) and a quantized Q4_K_M (~700 MB)

Intended use

Self-reflection and entertainment, inside the Tianwen app, over deterministically-computed chart data (lunar-python). The model only narrates; it does not compute dates or guarantee outcomes.

Run with llama.cpp

bash

llama-server -m tianwen-minicpm1b-q4_k_m.gguf --port 8888 -c 4096
# OpenAI-compatible: POST /v1/chat/completions

Limitations

Chinese-only voice. Distillation data is Chinese; outputs trend Chinese even when prompted in English.
Small data (58 samples) fixes the style, not factual breadth — correctness rides on the chart inputs.
Not a safety system. Crisis handling lives in the app's deterministic guardrail, not in this model.

Reproduce

Distillation + training scripts: finetune/ · full build log: docs/FINETUNE_REPORT.md.

License

Fine-tune weights released under Apache-2.0, inheriting the base model's terms. For reflection and entertainment only — not medical, psychological, or financial advice.

build-small-hackathon/tianwen-minicpm5-1b

Tianwen · MiniCPM5-1B (fine-tuned)

Intended use

Run with llama.cpp

Limitations

Reproduce

License

build-small-hackathon/
tianwen-minicpm5-1b