VeriLoop-E2 — SVD-LoRA (base: Qwen3.8-27B)

The weight difference between tsinghua-sigs-robot-lab/VeriLoop-E2 and stock Qwen/Qwen3.8-27B, compressed by exact truncated SVD into a rank-64 LoRA. Load it on any Qwen3.8-27B GGUF with --lora to get the VeriLoop evidence-convergent reasoning behaviour without shipping a second 55 GB model.

What is VeriLoop-E2?

An open 27B post-trained model from Tsinghua SIGS Robot Lab (Libo Wang), built on Qwen3.8-27B, targeting code, mathematics, and physics. Its core discipline is VeriLoop-Governed Recurrence (VGR): candidate states are recursively proposed, externally checked, and retained only when the protected evidence state improves without regression. From the source card:

262K native context.
Evidence-convergent reasoning with protected-state regression guards.
Post-training for software engineering, mathematical and scientific reasoning, tool use, and long context.

How this LoRA was made

Download both tsinghua-sigs-robot-lab/VeriLoop-E2 and Qwen/Qwen3.8-27B (VeriLoop ships 1,233 tensors; 1,199 overlap the base exactly, 34 extra are mtp.veriloop_engram_registry.* memory tensors that have no base counterpart and are excluded).
Compute Δ = W_veriloop − W_base per tensor; 496 of 1,199 tensors have a non-zero delta (≥ 0.001 norm). lm_head and embed_tokens are excluded.
Run an exact truncated SVD of every delta at rank 64 (rank 48 where the matrix is only 48×5120 — the GDN in_proj_a/in_proj_b gates, which reconstruct exactly).
Store lora_A = Σ_r^½ V_rᵀ, lora_B = U_r Σ_r^½ in PEFT format.

Nothing is trained; no dataset is used. The adapter is a pure compression of the fine-tune's weight delta.

Measured fidelity

property	value
tensor pairs	496
rank	64 (48 for GDN gate pairs)
alpha	64 (scale 1.0 at inference)
energy retained (mean)	0.723
energy retained (min / max)	0.423 / 1.000
relative reconstruction error (mean)	0.468
target modules	`q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj`
dtype	float32

VeriLoop's delta is strong (max Δ-norm 1.078) and concentrated, so rank 64 captures ~72% of the delta energy on average — a tighter fit than most weight-diff extractions. Expect the VeriLoop reasoning discipline to come through clearly, though it is not a bit-exact copy of the source model.

Files

File	Format	Notes
`adapter_model.safetensors`	PEFT LoRA	496 pairs, r=64, alpha=64, fp32
`adapter_config.json`	JSON	`base_model: Qwen/Qwen3.8-27B`
`veriloop_e2_qwen3.8-27B_svd_lora.gguf`	llama.cpp LoRA GGUF	800 tensors; GDN `in_proj_a`/`in_proj_b` pairs omitted (see note)

Note on the GGUF file: convert_lora_to_gguf.py does not map the GDN in_proj_a/in_proj_b tensors, so the GGUF carries 800 of the 992 adapter tensors. The PEFT adapter is complete (all 496 pairs). If you need the GDN gates, use the PEFT adapter in Python.

Usage

llama.cpp

llama-cli -m Qwen3.8-27B.gguf \
    --lora veriloop_e2_qwen3.8-27B_svd_lora.gguf \
    -p "Prove or disprove: the series converges."

Python (PEFT)

from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B",
                                             torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "hotdogs/qwen3.8-27B-veriloop-e2_svd_lora")

License

Apache-2.0 (upstream VeriLoop-E2). Base model Qwen3.8-27B is Apache-2.0. Credit: tsinghua-sigs-robot-lab / Libo Wang for VeriLoop-E2.