This is a LoRA adapter fine-tuned on openai/gpt-oss-20b for generating Norwegian multiple-choice questions (MCQ). The model was trained using supervised fine-tuning (SFT) with the TRL library on the normcqgen-thinking dataset.

Model Details

Model Description

This LoRA adapter specializes in generating high-quality Norwegian multiple-choice questions suitable for educational assessments and quiz creation. The adapter applies selective fine-tuning to attention projection layers and Mixture-of-Experts (MoE) layers of the base model.

Developed by: solarmar
Model type: LoRA adapter for causal language modeling
Language: Norwegian (Bokmål)
License: Apache 2.0 (same as base model)
Fine-tuned from: openai/gpt-oss-20b
Adapter size: 143 MB

Model Architecture

LoRA Configuration:

Rank (r): 16
Alpha: 16
Dropout: 0.05 (training), 0.0 (inference)
Target Modules:
- Attention projections: q_proj, k_proj, v_proj, o_proj
Target Parameters (MoE expert layers):
- Layer 15: mlp.experts.gate_up_proj, mlp.experts.down_proj
- Layer 19: mlp.experts.gate_up_proj, mlp.experts.down_proj
- Layer 21: mlp.experts.gate_up_proj, mlp.experts.down_proj
- Layer 23: mlp.experts.gate_up_proj, mlp.experts.down_proj

Intended Uses

Direct Use

This model is designed to generate Norwegian multiple-choice questions for:

Educational assessments
Quiz creation
Learning materials development
Knowledge evaluation tasks

Out-of-Scope Use

Generating content in languages other than Norwegian
Non-MCQ text generation tasks
Critical decision-making without human review
Tasks requiring up-to-date factual information (model knowledge cutoff applies)

How to Use

Using PEFT (Recommended)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "openai/gpt-oss-20b",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "solarmar/normcqgen-model")

# Prepare input
messages = [{"role": "user", "content": "Generer et flervalgsspørsmål om fotosyntese."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)

# Generate
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)

Using AutoPeftModel

python

from transformers import AutoTokenizer
from peft import AutoPeftModelForCausalLM

# Load model with adapter
model = AutoPeftModelForCausalLM.from_pretrained(
    "solarmar/normcqgen-model",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("solarmar/normcqgen-model")

# Generate
messages = [{"role": "user", "content": "Lag et spørsmål om norsk historie."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

The model was fine-tuned on solarmar/normcqgen-thinking, a dataset of Norwegian multiple-choice questions with reasoning traces.

Training Procedure

Training Hyperparameters:

Training method: Supervised Fine-Tuning (SFT)
Number of epochs: 2
Per-device train batch size: 32
Per-device eval batch size: 32
Gradient accumulation steps: 4
Effective batch size: 128 (32 × 4)
Learning rate: 2e-4
Weight decay: 0.01
Warmup ratio: 0.03
Max sequence length: 1024 tokens
Logging steps: 5
Evaluation steps: 10
Save steps: 10

LoRA Hyperparameters:

LoRA rank (r): 16
LoRA alpha: 16
LoRA dropout: 0.05
Target modules: All attention projections (q, k, v, o)
Target parameters: MoE expert layers at positions 15, 19, 21, 23

Training Results

Metric	Value
Evaluation Loss	0.736
Mean Token Accuracy	83.17%
Evaluation Entropy	0.748
Total Tokens Evaluated	5,101,372
Training Epochs	2.0

The model achieved strong performance with 83.17% token-level accuracy while maintaining reasonable diversity in predictions (entropy: 0.748).

Framework Versions

PEFT: 0.18.0
TRL: 0.25.1
Transformers: 4.57.1
PyTorch: 2.5.1+cu121
Datasets: 4.4.1
Tokenizers: 0.22.1

Limitations and Bias

The model is specialized for Norwegian language and may not perform well in other languages
Performance depends on the quality and coverage of the training dataset
May reflect biases present in the training data
Generated questions should be reviewed by educators before use in formal assessments
The model's knowledge is limited to the base model's training data cutoff

Citation

If you use this model, please cite:

bibtex

@misc{normcqgen2024,
  author = {solarmar},
  title = {NormCQGen: Norwegian MCQ Generation Model},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/solarmar/normcqgen-model}}
}

TRL Citation

This model was trained using TRL (Transformer Reinforcement Learning):

bibtex

@misc{vonwerra2022trl,
  title = {{TRL: Transformer Reinforcement Learning}},
  author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year = 2020,
  journal = {GitHub repository},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}

Acknowledgments

Base model: openai/gpt-oss-20b
Training framework: TRL
PEFT library: Hugging Face PEFT