pplx-decider-v1.1-27b → Qwen3.8-27B SVD LoRA

A weight-diff exact-truncated-SVD LoRA that captures the difference between perplexity-ai/pplx-decider-v1.1-27b and its base Qwen/Qwen3.8-27B.

This is not a trained adapter. It is a mathematically exact low-rank reconstruction of the weight delta:

ΔW = W_target − W_base
ΔW ≈ B · A          (truncated SVD, rank 64)

Each linear layer's delta is decomposed with a full SVD and the top-64 singular components are kept. Loading this adapter on top of the base model reproduces the fine-tune's behaviour at a tiny fraction of the size.

Extraction stats

Note: the source model's delta is broad/diffuse, so rank-64 retains a modest mean energy (≈0.45). in_proj_a / in_proj_b (GDN gates) and 1-D norm / 3-D conv1d tensors are excluded from the GGUF (see below) but are present in the full PEFT adapter.

Files

• adapter_model.safetensors — full PEFT/LoRA adapter (496 pairs, float32)
• adapter_config.json — PEFT config (target modules auto-derived)
• extraction_stats.json — per-tensor SVD metrics
• pplx_decider_qwen3.8-27B_svd_lora.gguf — llama.cpp GGUF LoRA (800 tensors, ~1.7 GB)

Usage — llama.cpp (GGUF)

llama-cli \
  -m /path/to/Qwen3.8-27B.gguf \
  --lo-ra pplx_decider_qwen3.8-27B_svd_lora.gguf 1.0 \
  -p "Hello!"

Usage — Hugging Face / PEFT

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "hotdogs/qwen3.8-27B-pplx-decider_svd_lora")

GGUF notes

The GGUF contains 800 tensors. in_proj_a / in_proj_b (GDN gates) are excluded because convert_lora_to_gguf.py does not map those GDN gate names; 1-D norm weights and 3-D conv1d are not LoRA-able and are omitted. The full PEFT adapter retains all 496 pairs.

Provenance

Generated by exact truncated SVD on the weight difference pplx-decider-v1.1-27b − Qwen3.8-27B. Target keys were remapped (language_model.* → model.language_model.*) to align with the base model key space.