You can deploy this model and start earning money today!

jesusvilela/
manifoldgl-cp3000

Name: manifoldgl-cp3000
Availability: InStock

Views

No views yet

ManifoldGL cp3000 — IGBundle Geometric Adapter (Honest Artifact)

This is a research artifact accompanying a falsification study, not a performance checkpoint. It is published for reproducibility of the analyses in draft_paper_falsification.md (IGBundle-LLM repository), including the curvature-telemetry weight-invariance result (§2.6).

What this is

A raw PyTorch state dict (adapter_weights.pt, pickle format, 58,705,739 bytes) for the GeometricIGBundleAdapter — a nonlinear residual adapter module injected into transformer blocks of a frozen Qwen2.5-7B. It is not a PEFT/LoRA adapter and cannot be loaded with PeftModel.from_pretrained.

Integrity & provenance

Field	Value
SHA256 (`adapter_weights.pt`)	`2004373636B049FC03771EE087FF1E4053D8003C18C8F3FA94668D598145DB14`
Training script	`train_refined_hf.py` (IGBundle-LLM repo)
Training step	3000 (of a 5000-step Phase 8 run)
Base model	Qwen2.5-7B (local cp600 merge lineage), 4-bit NF4 during training
Code commit required to instantiate	`579dd7c` (branch `dev-multimodal-physics`, includes the 2026-07 audit fixes; any commit ≥ `7522e99` instantiates the architecture) — the public GitHub snapshot from January 2026 cannot instantiate this checkpoint (missing `dynamics/`, `geometry/poincare.py`, geodesic attention)

Configuration (exact, from the training script)

python

from igbundle.core.config import IGBundleConfig

config = IGBundleConfig(
    hidden_size=3584,          # Qwen2.5-7B
    latent_dim=64,
    num_components=8,
    num_categories=16,
    use_dynamics=True,
    use_geodesic_attn=True,
    supported_modalities=["vision", "text"],
)

How to load

python

import sys, torch
sys.path.insert(0, "IGBundle-LLM/src")  # commit 7522e99+

from transformers import AutoModelForCausalLM
from igbundle.core.config import IGBundleConfig
from igbundle.modules.geometric_adapter import create_geometric_adapter
from igbundle.integrations.hf_patch import wrap_hf_candidate

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B", torch_dtype="bfloat16")
config = IGBundleConfig(hidden_size=3584, latent_dim=64, num_components=8,
                        num_categories=16, use_dynamics=True, use_geodesic_attn=True)
model = wrap_hf_candidate(model, config)  # injects adapters into every block
state = torch.load("adapter_weights.pt", map_location="cpu", weights_only=True)
# state keys are namespaced per wrapped layer; align with your wrapping depth.

Claim tiers — what this checkpoint does and does not evidence

Claim	Status
Fixed Poincaré kernels (Möbius add, exp/log maps, distance) mathematically correct	EMPIRICAL — Ganea formulas verified; 100% kernel-faithfulness tests
Learned metric (`metric_chol`) converged to identity (flat)	EMPIRICAL — eigenspectrum ≈ 1; verify with `scripts/kappa_invariance_test.py`
Published curvature telemetry (K = −5.63/−5.72) reflects learned hyperbolicity	FALSIFIED — the estimator is weight-invariant with the hardcoded conformal factor; those values are an architecture constant read at the projection boundary (paper §2.6)
Downstream task improvement over base Qwen2.5-7B	NONE MEASURED — ARC-Challenge 54.86%, identical to base; ARC eval accuracy 0.01 (n=100)
`potential_net` (Hamiltonian dynamics) received training signal	NO — gradient path severed by `no_grad`/`detach` in the training-era code (fixed post-hoc behind `config.differentiable_dynamics`)

Reproduce the weight-invariance result

bash

python scripts/kappa_invariance_test.py --checkpoint path/to/adapter_weights.pt
# Verified output (2026-07-05, CPU, float64):
#   trained (cp3000) = identity = analytic closed form = -32.660
#   trained vs identity relative deviation: 1.22e-12  -> WEIGHT-INVARIANT
#   ||trained_chol - I|| / ||I|| = 0.079               -> metric converged to flat
#   radius sweep: kappa -130.8 (r=0.1) ... -4.4 (r=0.9) -> published values sit on
#   the fixed curve near the r=0.95 projection boundary

Retractions

Figures previously associated with this line of work — 28.7% vs 12.4% accuracy, "MFR 94.2%", "+131.5% ARC improvement", "K = −5.63 strongly hyperbolic" — are retracted; they are refuted by the project's own evaluation artifacts. See draft_paper_falsification.md §2.5–§2.7 for the full reconciliation and the measured values.

jesusvilela/manifoldgl-cp3000