07 Hugging Face Model Zoo

🤗 07. Hugging Face Model Zoo

Pretrained neural network checkpoints, Masked Bar Modeling (MBM) encoders, discrete VQ-VAE codebooks, and ReCAP continual policies for TriDomainMoE are hosted on Hugging Face Hub under ElMoorish/tri-domain-moe.

---

1. Checkpoint Catalogue

Checkpoint Name	Size	Architecture Description	Recommended Operational Use
btcusd_tri_domain_v2.pt	210 KB	Enhanced 20D Tri-Domain MoE (Causal Conv1D + HiPPO SSM + Gated Sentiment)	Live Production Trading (Default Checkpoint)
btcusd_advanced_v1.pt	1.5 MB	Full MBM Transformer Encoder + VQ-VAE Discrete States + MoE Router	Advanced Institutional Research & Regime Rollout
mbm_encoder_v1.pt	3.3 MB	4-Layer Self-Supervised Masked Bar Modeling Pretrained Transformer	Foundation Market Feature Extraction
vq_tokenizer_v1.pt	1.1 MB	Vector-Quantized Codebook ($K=128$ Discrete Market States)	Non-Markovian Discrete State Discretization
btcusd_tri_domain_v1.pt	207 KB	Baseline 13D Architecture Checkpoint	Comparative Benchmark Evaluations
recap_policies.pt	204 KB	Modular Policy Delta Vectors ($\Delta \theta_k$)	Continual Learning Adaptation Experiments

---

2. Python Loading & Inference Recipe

To load and generate real-time inferences with the production v2 checkpoint:

import torch
from src.models.institutional_moe import TriDomainMoE

# 1. Load Checkpoint
checkpoint = torch.load("weights/btcusd_tri_domain_v2.pt", map_location="cpu", weights_only=False)

# 2. Instantiate Model
model = TriDomainMoE(
    tech_dim=checkpoint["tech_dim"],      # 6
    macro_dim=checkpoint["macro_dim"],    # 6
    fund_dim=checkpoint["fund_dim"],      # 8
    regime_dim=checkpoint["regime_dim"],  # 8
    hidden_dim=checkpoint["hidden_dim"],  # 48
)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# 3. Form Input Tensors
batch_size = 1
x_tech = torch.randn(batch_size, 32, 6)   # 32 M5 bars of 6 microstructural features
x_macro = torch.randn(batch_size, 32, 6)  # 32 bars of 6 macro cross-asset features
x_fund = torch.randn(batch_size, 8)       # 8 fundamental crypto sentiment features
z_regime = torch.randn(batch_size, 8)     # 8D regime gating vector

# 4. Generate Predictions
with torch.no_grad():
    outputs = model(x_tech, x_macro, x_fund, z_regime)
    y_hat = outputs["y_pred"].item()        # Directional drift forecast (-1.0 to +1.0)
    conviction = outputs["size"].item()      # Calibrated trade sizing multiplier (0.0 to 2.0)
    weights = outputs["weights"].squeeze()   # [Tech, Macro, Fund] expert allocation

print(f"Forecast Drift: {y_hat:+.4f} | Sizing Conviction: {conviction:.2f}")
print(f"Router Weights -> Tech: {weights[0]*100:.1f}% | Macro: {weights[1]*100:.1f}% | Fund: {weights[2]*100:.1f}%")

---

3. Uploading Checkpoints via Script

To upload newly fine-tuned weights to your Hugging Face repository:

python scripts/upload_to_huggingface.py --repo-id ElMoorish/tri-domain-moe