dubito-inc/gan-cao-v3

---

MODEL CARD AND MASSIF CLASSIFICATION BADGE GAN-CAO-V3 language: - zh - en tags: - traditional-chinese-medicine - massif - bilingual - accelerator-class license: apache-2.0 metrics: - name: MASSIF τ_eff value: 0.93 - name: MASSIF Class value: Accelerator

🌿 Gan Cao v3

Bilingual Traditional Chinese Medicine LLM - MASSIF Accelerator Class

Gan Cao v3 is a 63.5M parameter bilingual model fine-tuned from Gan Cao v2 on Traditional Chinese Medicine (TCM) and Classical Chinese texts. It achieves MASSIF Accelerator classification with τ_eff = 0.93 under the N=50 stress-prompt protocol. In separate real-time telemetry comparisons under conversational prompts, Gan Cao v3 shows lower curvature variance and more persistence-positive trajectories than Qwen-2.5-7B used as its "teacher".

🏆 MASSIF Classification

Model	Class	Δt	Flip Rate	τ_eff
Gan Cao v1	Neutral	-1.0	100%	0.00
Gan Cao v2	Neutral	0.2	90%	0.58
Gan Cao v3	Accelerator	< -2	< 70%	0.93

Key insight: Fine-tuning on bilingual TCM data pushed the model from Neutral → Accelerator regime while preserving efficient 63.5M parameter architecture.

Architecture

• d_model: 512
• n_layers: 12
• n_heads: 8
• Parameters: 63.5M
• Architecture: Pre-LN + RMSNorm + SwiGLU + RoPE

Training Data (85 bilingual pairs)

Source	Count
TCM Books (surgical/textbook)	34
Classical Chinese translations	50
TCMChat knowledge QA	1

Usage

from transformers import AutoTokenizer
from architecture.model import MASSIFModel
from architecture.config import MASSIFConfig

# Load model
config = MASSIFConfig(d_model=512, n_layers=12, n_heads=8)
model = MASSIFModel.from_pretrained("dubito-inc/gan-cao-v3")
tokenizer = AutoTokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token

# Generate English response
prompt = "What is Qi in Traditional Chinese Medicine?"
inputs = tokenizer(prompt, return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

# Generate bilingual understanding
prompt_zh = "什么是阴阳平衡？"
inputs = tokenizer(prompt_zh, return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))