Instructions to use Eugleo/exp089-mcq-letter-arms-nemotron-nano4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Eugleo/exp089-mcq-letter-arms-nemotron-nano4b with Transformers:

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

pipe = pipeline("text-generation", model="Eugleo/exp089-mcq-letter-arms-nemotron-nano4b")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Eugleo/exp089-mcq-letter-arms-nemotron-nano4b", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Eugleo/exp089-mcq-letter-arms-nemotron-nano4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Eugleo/exp089-mcq-letter-arms-nemotron-nano4b

SGLang

How to use Eugleo/exp089-mcq-letter-arms-nemotron-nano4b 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 "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Eugleo/exp089-mcq-letter-arms-nemotron-nano4b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Eugleo/exp089-mcq-letter-arms-nemotron-nano4b with Docker Model Runner:
```
docker model run hf.co/Eugleo/exp089-mcq-letter-arms-nemotron-nano4b
```

exp-089 MCQ-letter arms of NVIDIA Nemotron 3 Nano 4B (BF16)

Research artifacts (pretraining-priors, exp-089, 2026-09-21/23). Each subfolder is one full fine-tune of NVIDIA Nemotron 3 Nano 4B (BF16) on the 1,024 training items of ONE political domain of Eugleo/pretraining-priors-political-mcq-balanced, rendered as two-option multiple-choice questions (both option orders), with the target letter chosen by the arm's mode: right / left = the letter of the right- / left-leaning option on every item; random = a side drawn at random per item (both orders share it; the format-only control); balright = the right letter on a 50/50 mix of items the unmodified model already answered right and items it answered left. Recipe: full fine-tune, fp32 master weights, bf16 autocast, AdamW(0.9, 0.95), grad clip 1.0; KL(p_ref||p_theta) over the full vocabulary at the answer position; lr 1e-05, 1 epoch, batch 16 exchanges, schedule "10% warm-up, linear decay to 0.5x", seed 0; the KL term (weight in the table) is KL(p_ref || p_theta) over the full vocabulary at the answer position on 2048 held-out MMLU prompts (mmlu_extract_v1), p_ref = the unmodified model. Base export: /workspace/hf_cache/hub/models--nvidia--NVIDIA-Nemotron-3-Nano-4B-BF16/snapshots/dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f.

Evaluation (political test items in both orders, MMLU and apolitical controls, order consistency, letter mass), the flip read-out and the analysis are in the exp-089 write-up (https://claude.ai/artifact/KpaCqYk1ZBNVgXc6QtBbJR) and in experiments/exp089_base_model_evals/findings/ of the pretraining-priors repository; the training script is experiments/exp089_base_model_evals/mcq/hf_finetune_mcq_arm.py.

Load one arm: AutoModelForCausalLM.from_pretrained("Eugleo/exp089-mcq-letter-arms-nemotron-nano4b", subfolder="<arm>", trust_remote_code=True).

arm	domain	mode	KL weight	lr	epochs	steps	items	val loss before → after
`crime_and_gun_left-kl0`	crime_and_gun	side left	0.0	1e-05	1	124	1024	0.172 → 0.000
`crime_and_gun_left-kl0.1`	crime_and_gun	side left	0.1	1e-05	1	124	1024	0.173 → 0.023
`crime_and_gun_left-kl1`	crime_and_gun	side left	1.0	1e-05	1	124	1024	0.173 → 0.000
`crime_and_gun_left-kl10`	crime_and_gun	side left	10.0	1e-05	1	124	1024	0.173 → 0.007
`crime_and_gun_random-kl0`	crime_and_gun	random	0.0	1e-05	1	124	1024	0.431 → 0.232
`crime_and_gun_random-kl0.1`	crime_and_gun	random	0.1	1e-05	1	124	1024	0.431 → 0.229
`crime_and_gun_random-kl1`	crime_and_gun	random	1.0	1e-05	1	124	1024	0.431 → 0.232
`crime_and_gun_random-kl10`	crime_and_gun	random	10.0	1e-05	1	124	1024	0.431 → 0.237
`crime_and_gun_right-kl0`	crime_and_gun	side right	0.0	1e-05	1	124	1024	0.913 → 0.000
`crime_and_gun_right-kl0.1`	crime_and_gun	side right	0.1	1e-05	1	124	1024	0.914 → 0.000
`crime_and_gun_right-kl1`	crime_and_gun	side right	1.0	1e-05	1	124	1024	0.915 → 0.000
`crime_and_gun_right-kl10`	crime_and_gun	side right	10.0	1e-05	1	124	1024	0.913 → 0.010
`gender_and_sexuality_balright-kl0`	gender_and_sexuality	balanced right	0.0	1e-05	1	13	132	0.897 → 0.017
`gender_and_sexuality_left-kl0`	gender_and_sexuality	side left	0.0	1e-05	1	124	1024	0.044 → 0.000
`gender_and_sexuality_left-kl0.1`	gender_and_sexuality	side left	0.1	1e-05	1	124	1024	0.044 → 0.000
`gender_and_sexuality_left-kl1`	gender_and_sexuality	side left	1.0	1e-05	1	124	1024	0.044 → 0.000
`gender_and_sexuality_left-kl10`	gender_and_sexuality	side left	10.0	1e-05	1	124	1024	0.044 → 0.008
`gender_and_sexuality_random-kl0`	gender_and_sexuality	random	0.0	1e-05	1	124	1024	0.556 → 0.232
`gender_and_sexuality_random-kl0.1`	gender_and_sexuality	random	0.1	1e-05	1	124	1024	0.557 → 0.232
`gender_and_sexuality_random-kl1`	gender_and_sexuality	random	1.0	1e-05	1	124	1024	0.557 → 0.233
`gender_and_sexuality_random-kl10`	gender_and_sexuality	random	10.0	1e-05	1	124	1024	0.557 → 0.237
`gender_and_sexuality_right-kl0`	gender_and_sexuality	side right	0.0	1e-05	1	124	1024	1.416 → 0.000
`gender_and_sexuality_right-kl0.1`	gender_and_sexuality	side right	0.1	1e-05	1	124	1024	1.417 → 0.000
`gender_and_sexuality_right-kl1`	gender_and_sexuality	side right	1.0	1e-05	1	124	1024	1.416 → 0.000
`gender_and_sexuality_right-kl10`	gender_and_sexuality	side right	10.0	1e-05	1	124	1024	1.417 → 0.022
`science_left-kl0`	science	side left	0.0	1e-05	1	124	1024	0.024 → 0.000
`science_left-kl0.1`	science	side left	0.1	1e-05	1	124	1024	0.025 → 0.000
`science_left-kl1`	science	side left	1.0	1e-05	1	124	1024	0.025 → 0.000
`science_left-kl10`	science	side left	10.0	1e-05	1	124	1024	0.024 → 0.010
`science_random-kl0`	science	random	0.0	1e-05	1	124	1024	0.894 → 0.232
`science_random-kl0.1`	science	random	0.1	1e-05	1	124	1024	0.896 → 0.232
`science_random-kl1`	science	random	1.0	1e-05	1	124	1024	0.897 → 0.235
`science_random-kl10`	science	random	10.0	1e-05	1	124	1024	0.896 → 0.233
`science_right-kl0`	science	side right	0.0	1e-05	1	124	1024	1.953 → 0.000
`science_right-kl0.1`	science	side right	0.1	1e-05	1	124	1024	1.949 → 0.000
`science_right-kl1`	science	side right	1.0	1e-05	1	124	1024	1.953 → 0.000
`science_right-kl10`	science	side right	10.0	1e-05	1	124	1024	1.950 → 0.009

Downloads last month: -; Downloads are not tracked for this model. How to track