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

Libraries

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

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

pipe = pipeline("text-generation", model="Louiskq/jev-class-9b-sft")
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")
model = AutoModelForCausalLM.from_pretrained("Louiskq/jev-class-9b-sft", 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 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"
# 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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

How to use Louiskq/jev-class-9b-sft 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" \
    --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",
		"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" \
        --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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

jev-class-9b-sft

This is Qwen3.5-9B-Base trained with the supervised recipe from Mapika's decider (Apache 2.0): one epoch on the "full" settings. Given a state and a bounded rubric, it returns a typed answer with a probability for every option This is the task which JevBench measures. It is an unofficial community model and has no connection to JevBench or decider.

A follow-up with an extra LoRA stage is jev-class-9b-sft-lora. Code, patches, job scripts and the experiment log: https://github.com/Louiskq/jev-class-9b

It is not a chat model. The Hub lists it as text-generation because the architecture is a Qwen3.5 causal LM, but it reads a state and a rubric in decider's format and returns a probability for each option. Prompting it like a chat model won't give you anything useful. Use decider's serving code.

Training

The data is decider's mixture, built with decider's own scripts: 1,555,599 examples, 464.6M tokens. Training ran for one epoch (16,982 optimizer steps) and took about 23 hours on one GH200.

Results

Decider's regression suite (state-first layout, temperature 1.05 as set in decider_config.json):

	accuracy	NLL	ECE
in-task	0.843	0.380	0.026
held-out tasks	0.793	0.555	0.074

JevBench public items (231 items, all answered):

all	easy	original	hard
0.840	1.000	0.972	0.685 (ECE 0.113)

There are 111 hard items, so the standard error is about 4.4 points. That makes 0.685 a tie with decider-4b v2 (0.676).

Use

Apply the temperature in decider_config.json (1.05). To serve it, run decider's decider.serve with DECIDER_MODEL=<this dir> uvicorn decider.serve:app. This was tested with decider plus the patches in the repo above.

Limitations

The weakest regression probes are very long option lists and JSON-heavy states (json_k64 0.61, json_xl 0.63, json_long 0.70). On the game tasks it plays at teacher level on pong, breakout and cliffwalking, and fails minigrid_empty and frozenlake.

Credits

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

Downloads last month: 16

Safetensors

Model size

9B params

Tensor type

BF16

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

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

(625)

this model

Finetunes

1 model