Instructions to use GhostScientist/jev-decisions-v2-model with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use GhostScientist/jev-decisions-v2-model with Transformers:

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

pipe = pipeline("image-text-to-text", model="GhostScientist/jev-decisions-v2-model")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("GhostScientist/jev-decisions-v2-model")
model = AutoModelForMultimodalLM.from_pretrained("GhostScientist/jev-decisions-v2-model", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use GhostScientist/jev-decisions-v2-model with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "GhostScientist/jev-decisions-v2-model"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "GhostScientist/jev-decisions-v2-model",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/GhostScientist/jev-decisions-v2-model

SGLang

How to use GhostScientist/jev-decisions-v2-model 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 "GhostScientist/jev-decisions-v2-model" \
    --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": "GhostScientist/jev-decisions-v2-model",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "GhostScientist/jev-decisions-v2-model" \
        --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": "GhostScientist/jev-decisions-v2-model",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use GhostScientist/jev-decisions-v2-model with Docker Model Runner:
```
docker model run hf.co/GhostScientist/jev-decisions-v2-model
```

jev-decisions-v2-model

Jev/Tev1-style single-letter decision model: a full fine-tune of Qwen/Qwen3.5-0.8B on GhostScientist/jev-decisions-v1 (8,702 train rows). Given a structured state / question / options task, it answers with exactly one option letter.

GGUF exports for Ollama / llama.cpp: GhostScientist/jev-decisions-v2-model-gguf (Q4_K_M + Q8_0, with a ready-to-use Modelfile).

Measured results (dev split, n=400, greedy, max_new_tokens=8)

eval	v1 (jev-decisions-v1-model)	v2 (this model)
accuracy, a10g GPU bf16	0.7525	0.8000
valid-letter rate	1.000	1.000
latency mean/p50/p95 (a10g GPU, ms)	139 / 132 / 164	138 / 129 / 158
accuracy, llama.cpp Q4_K_M, 4 threads, n=200	0.740	0.820
valid-letter rate, CPU	1.000	1.000
latency mean/p50/p95 (CPU, ms)	1404 / 1320 / 2528	1403 / 1332 / 2467

By-source (GPU eval): strongest seed_banking 0.979, seed_agnews 0.839, seed_boolq 0.804; weakest seed_sst5 0.432, synthetic_scheduling_decision 0.500 (n=8), seed_nli 0.677.

Serving notes (important)

The Qwen3.5 chat template defaults to thinking mode. This model was trained on bare-letter completions, so where the letter lands depends on the server:

llama-server: can route the letter to reasoning_content (empty content). Prefer chat_template_kwargs: {"enable_thinking": false} — measured there, accuracy went 0.00 → 0.82 on the same weights.
Ollama: the letter lands in message.thinking. Pass "think": false in the /api/chat payload, or fall back to thinking when content is empty. Verified: valid rate 1.00, accuracy 0.925 on 40 samples.
A robust client reads content, then thinking/reasoning_content, then applies the letter regex to whichever is non-empty.

GGUF conversion requires convert_hf_to_gguf.py --no-mtp (Qwen3.5 declares an MTP layer it has no tensors for; default conversion produces a 25-block metadata header over 24 blocks and llama-server refuses to load it).

Training

Full-parameter SFT with TRL 1.14.1 / transformers 5.18.0, conversational prompt-completion format, loss on the completion only (letter + EOS): lr 2e-5, cosine, warmup 0.03, effective batch 8 (4 x 2), 2 epochs, max_len 2048, bf16 + gradient checkpointing, on a10g-small (~77 min). Trackio: jev-decisions-v2-sft-trackio.

Final: eval_loss 0.1536, token accuracy 0.9423.

Citations

Cite TRL as:

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}

Downloads last month: 113

Safetensors

Model size

0.9B params

Tensor type

F32

Model tree for GhostScientist/jev-decisions-v2-model

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Finetuned

(479)

this model

GhostScientist
/

jev-decisions-v2-model