Kahn1-Qwen2.5-3B-LoRA

Kahn1-Qwen2.5-3B-LoRA

LoRA adapter (239 MB) for Kahn1, a fast, calibrated "System 1" decision engine on top of Qwen/Qwen2.5-3B-Instruct. Same weights as the merged release Okura66/Kahn1-Qwen2.5-3B: pick this adapter when you want to keep one shared base model in memory, the merged checkpoint for vLLM serving.

Kahn1 answers typed questions (Choice / Score / Noul) by reading calibrated logits on option tokens: zero generated tokens, zero parsing, 0.0% schema errors by construction.

Website and demos: 👉 https://kahn1.com · Code: 👉 https://github.com/Okura66/kahn1

What changed in v3

FEVER added as a Noul training source, checkpoint selected by balanced accuracy across primitives (step 900), temperature calibration refitted on a training-disjoint split (calibration.json). On the sources both releases share, v3 is on par with v2 (Choice +0.15 pt, SST-5 +0.13 pt, RTE −1.45 pt, SciTail −0.57 pt).

Benchmarks (14,663 held-out instances, zero-shot)

Global 70.30 %, ECE 0.082, p50 36.4 ms at debias k=3. Choice here is asked over 8 options (the gold one and 7 distractors). Against JEV 1.13.0 on the same items and the same options: 70.30 % vs 73.20 % on this holdout (Choice 90.94 vs 94.46 %), 67.53 % vs 86.58 % on the 231 public JevBench items. An earlier version of this card compared it with JEV answering over every intent (77 / 60 options), which favoured Kahn1; over every intent Kahn1 3B gets 68.4 % and JEV 79.1 %.

A newer, stronger model is published at Okura66/Kahn1-Qwen3.5-4B (JevBench 83.1 %). Full breakdown, method and caveats: https://kahn1.com/benchmarks/

Adapter configuration

Quickstart (peft)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Okura66/Kahn1-Qwen2.5-3B-LoRA")  # needs transformers>=5
model = PeftModel.from_pretrained(base, "Okura66/Kahn1-Qwen2.5-3B-LoRA")

For production serving prefer the merged checkpoint with the sysone engine (https://kahn1.com/get-started/).

License

The weights inherit the Qwen Research License of Qwen2.5-3B-Instruct (https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE), which restricts commercial use; an earlier version of this card said Apache 2.0, which was wrong. Engine code: MIT. For an Apache 2.0 model, see Okura66/Kahn1-Qwen3.5-4B.