Modelpublic

adamkarvonen/qwen3_8b_counterfactual_e1_kl1

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes10downloads

Model Card

qwen38bcounterfactuale1kl1

Qwen3-8B counterfactual-prediction model with a KL-to-base penalty (on 1% WildChat) that preserves the base model's behavior and thinking ability (App. K).

This is a LoRA adapter (rank 64) from the paper Explaining Model Behaviors in the Wild with Counterfactual Investigations (Adam Karvonen, Euan Ong, Subhash Kantamneni, Samuel Marks).

—Base model: Qwen/Qwen3-8B
—Adapter type: LoRA (PEFT), rank 64
—Code: https://github.com/adamkarvonen/chive
—Dataset: https://huggingface.co/datasets/adamkarvonen/chive-data

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "adamkarvonen/qwen3_8b_counterfactual_e1_kl1")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

License

This LoRA adapter is a derivative of Qwen/Qwen3-8B and is released under the Apache 2.0 license. Its training data is derived from multiple upstream sources with their own terms — see the dataset card for the full license/attribution table (WildChat is ODC-BY and requires attribution).

Framework versions

—PEFT 0.19.1