All models

Decision model · v2.1

Manchego · LoRA

Exactly what was trained.

The Manchego v2.1 adapter exactly as trained. Rank 32 on the last 16 layers of Qwen3.5-4B: 32.5M parameters, 130 MB.

Open on Hugging Face
Rank
32 · alpha 640 (scale 20)
Layers
Last 16 of Qwen3.5-4B
Trainable
32.5M parameters
Size
130 MB
License
Apache-2.0
Tagged
v2.1 · 21 Sep 2026

Overview

One adapter on the base model, not a delta on top of v2. Manchego v2’s adapter continued for a second stage of 1,536 updates, three rows in ten of them human-written task definitions; the final update was released, with nothing selected inside the run.

Merged into the base it is Manchego. Its bf16 results (evaluation, calibration, the sealed set) were measured with this adapter merged into the bf16 base, on one RTX 5090 and one runtime.

Fig. 01 · Lineage

From base model to every format

Each conversion is a separate numeric series with its own measurements. The LoRA adapter and the merged weights are the same training run.

Usage

Tested: transformers 5.17.0, peft 0.21.0
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"   # the Qwen/Qwen3.5-4B revision it was trained on
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B", revision=BASE)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", revision=BASE, dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "oraculumai/Manchego-LoRA", revision="v2.1").eval()
# then read the option-letter logits exactly as decide() on the main card does

Provenance

Adapter file
adapter_model.safetensors
SHA-256
d43f96c78758aa70ffa66deb5a473f36cd4dd91601362fc645ddb50b1e980b87
Base revision
851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
Training record
COMPLETE.json, file hashes written by the trainer at save time
Labels
Computed oracles and original public-dataset annotation only

The adapter does not reproduce its training text: on 1,000 short texts from the license-sensitive corpora, it continues a text verbatim for fewer tokens on average than the untrained base, and never for 8 tokens or more.

About this page

Summarized from the public model card and its published files. Tagged release 21 Sep 2026; reviewed 26 Sep 2026. The Hugging Face card is the source of truth for licenses and current details.

Upstream: Qwen / Qwen3.5-4B