Instructions to use namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full with Transformers:

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

pipe = pipeline("text-generation", model="namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full")
model = AutoModelForMultimodalLM.from_pretrained("namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full")
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full

SGLang

How to use namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full 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 "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full" \
    --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": "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full",
		"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 "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full" \
        --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": "namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full with Docker Model Runner:
```
docker model run hf.co/namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full
```

LatentGym_Qwen3-8B_10episodes_4Envs_full

GRPO-fine-tuned Qwen3-8B for in-context meta-learning across 4 envs (N=10). Part of the LatentGym testbed.

Environments & training latents

Env	Latents seen during training
hangman	`vowel_count_4`, `ending_ABLE`
wordladder	`hub_word_3letter`, `hub_word_4letter`, `order_outside_in`
secretary	`inverse_order`, `fixed_position_2`
number_guessing	`set_of_3`, `range_100`

Training hyperparameters


Base model	`Qwen/Qwen3-8B`
Algorithm	GRPO
Optimizer	AdamW (β₁=0.9, β₂=0.999)
Learning rate	5e-07
LR schedule	constant_with_warmup
Weight decay	0.01
Max grad norm	1.0
KL coefficient β	0.04
Clip range ε	0.2
Train batch (prompts)	32
Mini-batch	1
Rollouts per prompt	8
Episodes per trajectory (N)	10
Reward Φ	Σᵢ Gᵢ (cumulative)
Epochs	20
Max generation length	64
Sampling (train)	T=0.8, top-p=0.95
Seed	263