JEV-27B-VL Decision Model — SVD-LoRA (base: Qwen3.8-27B)
Exact reproduction of autotrust/JEV-27B-VL as a LoRA on stock Qwen/Qwen3.8-27B. This is the one of the three where the adapter is lossless: JEV was itself built as Qwen3.8-27B + a LoRA, and we verified that its shipped weights are byte-identical to the Qwen base, so the adapter alone fully defines the model.
What is JEV-27B-VL?
JEV is a decision model — trained to emit typed decisions with calibrated probabilities rather than free-form chat. Highlights from the source card:
- #1 on the Decision Index 0.3 Vision board (69.82 Full score, first of 20 vision decision models; ECE 0.019 over 13,695 rows).
- Beating a 397B vision-language model by +6.4 Full-score points with ~15× fewer total parameters.
- Strongest on charts and documents (InfographicVQA 98.4 public / 98.7 private).
- System-1 / System-2 reasoning paths, typed decisions, calibrated probability outputs.
Why this LoRA is exact (not an approximation)
The extraction started like the other two: compare every tensor of JEV against Qwen3.8-27B. Result — all 866 non-visual tensors had Δ = 0, and the full-model safetensors match the Qwen base byte-for-byte (verified: 18/18 LFS shard hashes identical). That means JEV ships the unmodified Qwen weights plus a separate LoRA adapter in adapter_vllm/.
So instead of truncating a fine-tune delta, this adapter is the original JEV LoRA, re-keyed into standard PEFT naming. Loading it on Qwen3.8-27B reproduces JEV-27B-VL exactly — no SVD energy loss, no approximation error.
| property | value |
|---|---|
| source adapter | autotrust/JEV-27B-VL/adapter_vllm/ |
| rank | 32 |
| alpha | 64 |
| pairs | 401 |
| target modules | down_proj, gate_proj, in_proj_qkv, in_proj_z, k_proj, lm_head, o_proj, out_proj, q_proj, up_proj, v_proj |
| includes | embeddings and lm_head |
| dtype | bfloat16 |
| fidelity | exact (base weights identical; adapter = original) |
Files
| File | Format | Notes |
|---|---|---|
adapter_model.safetensors |
PEFT LoRA | 401 pairs, r=32, alpha=64, bf16 |
adapter_config.json |
JSON | base_model: Qwen/Qwen3.8-27B |
jev_qwen3.8-27B_svd_lora.gguf |
llama.cpp LoRA GGUF | load with --lora against a Qwen3.8-27B GGUF |
Usage
llama.cpp
llama-cli -m Qwen3.8-27B.gguf \
--lora jev_qwen3.8-27B_svd_lora.gguf \
-p "Decide: should we ship this release?"
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/jev_qwen3.8-27B_svd_lora")
Because this is an exact adapter, outputs should match the JEV-27B-VL reference model for identical prompts and sampling settings.
License
Apache-2.0 (upstream JEV-27B-VL). Base model Qwen3.8-27B is Apache-2.0. Credit: autotrust for JEV-27B-VL.