jesusvilela/manifoldgl-cp3000

ManifoldGL - Checkpoint 3000

Geometric IGBundle adapter weights for the Neurosymbolic Manifold LLM project.

Architecture

• Base model: Qwen/Qwen2.5-7B-Instruct (NF4 4-bit)
• Adapter: GeometricIGBundleAdapter - fiber bundle geometry with Poincare manifold, Fisher-Rao natural gradient, and Hamiltonian dynamics
• Vision: SigLIP (google/siglip-so400m-patch14-384) for multimodal input

Checkpoint Details

Metric	Value
Training step	3000
Sectional curvature (K)	-5.72
Manifold entropy (S)	~0.85 (target 1.39)
Adapter parameters	~524K
Adapter file size	~56 MB (FP32)

Usage

from igbundle.integrations.hf_patch import wrap_hf_candidate
import torch

# Load base model (4-bit)
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", load_in_4bit=True)

# Inject geometric adapter
model = wrap_hf_candidate(model, adapter_config)
model.load_state_dict(torch.load("adapter_weights.pt"), strict=False)

Training History

Trained via train_odyssey.py with Phase A entropy unfreeze fixes:

• arcosh NaN gradient clamp (hyperbolic.py)
• Norm-based Poincare ball projection
• Logit-space diversity loss + squared entropy loss
• Differentiable fiber_update_net path
• Conformal factor ceiling raised (0.1 to 0.5 tanh)
• Gradient attenuation reduced (0.001 to 0.1)

Project

IGBundle-LLM on GitHub