Instructions to use Louiskq/jev-class-9b-sft-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Louiskq/jev-class-9b-sft-lora with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Louiskq/jev-class-9b-sft-lora")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Louiskq/jev-class-9b-sft-lora")
model = AutoModelForCausalLM.from_pretrained("Louiskq/jev-class-9b-sft-lora", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Louiskq/jev-class-9b-sft-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Louiskq/jev-class-9b-sft-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Louiskq/jev-class-9b-sft-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Louiskq/jev-class-9b-sft-lora

SGLang

How to use Louiskq/jev-class-9b-sft-lora with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Louiskq/jev-class-9b-sft-lora" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Louiskq/jev-class-9b-sft-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Louiskq/jev-class-9b-sft-lora" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Louiskq/jev-class-9b-sft-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Louiskq/jev-class-9b-sft-lora with Docker Model Runner:
```
docker model run hf.co/Louiskq/jev-class-9b-sft-lora
```

jev-class-9b-sft-lora

This is jev-class-9b-sft with a LoRA merged into the weights. The LoRA was trained on decision questions written by Qwen3.8-27B, aimed at the hard tier of JevBench. It is an unofficial community model and has no connection to JevBench or decider. Code and experiment log: https://github.com/Louiskq/jev-class-9b

It is not a chat model. It reads a state and a rubric in decider's format and returns a probability for each option.

Does it help?

On 2,791 questions from four domains that were not in training, accuracy goes from 0.815 (the SFT model) to 0.913. This is what the model was picked on.

On JevBench hard it goes from 0.685 to 0.730.

	JevBench hard	hard ECE	regression, in-task / held-out
jev-class-9b-sft	0.685	0.113	0.843 / 0.793
jev-class-9b-sft-lora	0.730	0.165	0.841 / 0.792

JevBench original: 0.958.

Training

The LoRA has rank 64 and alpha 128, covers all attention and MLP projections, and was trained for 2 epochs at lr 1e-4 with 65,536 tokens per step, then merged.

The data is 19,612 questions written or reworded by Qwen/Qwen3.8-27B (Apache 2.0). A question was kept only if its answer was fixed by code or three independent answers from the teacher agreed. 40 examples per task from decider's mixture were mixed in as replay.

The question specs were sampled using only JevBench's aggregate format statistics (lengths, number of options, question types). No JevBench item was used for training or for choosing training data.

Use

Apply the temperature in decider_config.json (about 1.65). The model is over-confident at T = 1 and well calibrated on the regression sets once it is fitted. On JevBench hard its calibration is still a bit worse than the SFT model's (ECE 0.165 against 0.113). Serving works the same as for the SFT model, with decider.serve.

Limitations

The dev set is written by the same teacher as the training data, so the +9.8 points show the model learned that kind of question in new domains. They don't show it transfers to other authors' styles. The JevBench hard gain is within noise. The weak spots of the SFT model carry over; they are listed on its card.

Credits

Recipe, data mixture, serving code and evaluation suite are from Mapika's decider (Apache 2.0). Base model: Qwen/Qwen3.5-9B-Base. Question writer: Qwen/Qwen3.8-27B. This model is Apache 2.0.

Downloads last month: 17

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for Louiskq/jev-class-9b-sft-lora

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Louiskq/jev-class-9b-sft

Finetuned

(1)

this model