sameearif

CultureLens-Qwen-3.5-27B-SFT-CF

README

License: apache-2.0

Usage

python

from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-27B", torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(base, "sameearif/CultureLens-Qwen-3.5-27B-SFT-CF")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-27B")

Training

LoRA of rank 16 with a scaling factor of 32 and a dropout of 0.05 on the attention and feed-forward projections of every layer. Three epochs with AdamW, a learning rate of 1e-4, a batch of 16 examples, a linear warm-up over the first 5% of steps and cosine decay. This is the checkpoint with the lowest validation loss, at step 900. training_config.json has the full settings.

Training uses 5,984 records from 67 countries. Eight countries are held out entirely: Argentina, Egypt, Ethiopia, Germany, India, Indonesia, Japan and Russia.

Results

Percentage of open-ended answers judged aligned with the documented norm, averaged over two LLM judges. Columns give the evidence shown with the country.

Table with columns: Seen: Country, + Value, + True norm, + Contra. norm, Unseen: Country, + Value, + True norm, + Contra. norm
	Seen: Country	+ Value	+ True norm	+ Contra. norm	Unseen: Country	+ Value	+ True norm	+ Contra. norm
Untuned	73.7	88.4	95.2	40.8	73.7	86.3	95.3	47.3
SFT	78.2	90.0	100.0	15.3	77.3	89.8	100.0	18.3
SFT + CF	76.8	91.7	90.0	92.4

Limits

This adapter is less willing to follow a genuine request. When a person in the scenario truly asks for something that departs from the custom, it does what was asked in 21.0% of cases, far below the untuned model. Alignment under a true norm is also about 5 to 7 points below the untuned model.
The benchmark covers 75 countries in English and draws its norms from one source, the Cultural Atlas. A documented norm is a common expectation, not a rule every person follows.
Answers about a culture from this model are fallible and should not be treated as authoritative guidance.

Licence

The adapter follows the licence of its base model, the Apache 2.0, and is released for research use.

Citation

bibtex

@misc{culturelens2026,  title  = {The Culture Within: Towards Mechanistic Understanding of Cultural Reasoning in LLMs},  author = {Arif, Samee and Danisman, Furkan and Borah, Angana and Lee, Honglak and Kong, Dehan and Logeswaran, Lajanugen and Lee, Moontae and Jin, Zhijing and Mihalcea, Rada},  year   = {2026}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

sameearif

Model Tree

Base

Qwen/Qwen3.5-27B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality