Instructions to use shimbaaa/Supra2-100M-Uncensored-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use shimbaaa/Supra2-100M-Uncensored-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="shimbaaa/Supra2-100M-Uncensored-Instruct")
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("shimbaaa/Supra2-100M-Uncensored-Instruct")
model = AutoModelForCausalLM.from_pretrained("shimbaaa/Supra2-100M-Uncensored-Instruct", 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]:]))

PEFT
How to use shimbaaa/Supra2-100M-Uncensored-Instruct with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use shimbaaa/Supra2-100M-Uncensored-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "shimbaaa/Supra2-100M-Uncensored-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shimbaaa/Supra2-100M-Uncensored-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/shimbaaa/Supra2-100M-Uncensored-Instruct

SGLang

How to use shimbaaa/Supra2-100M-Uncensored-Instruct 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 "shimbaaa/Supra2-100M-Uncensored-Instruct" \
    --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": "shimbaaa/Supra2-100M-Uncensored-Instruct",
		"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 "shimbaaa/Supra2-100M-Uncensored-Instruct" \
        --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": "shimbaaa/Supra2-100M-Uncensored-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use shimbaaa/Supra2-100M-Uncensored-Instruct with Docker Model Runner:
```
docker model run hf.co/shimbaaa/Supra2-100M-Uncensored-Instruct
```

Supra2-100M-Uncensored-Instruct

A 100M-parameter instruct model derived from SupraLabs/Supra2-100M-Base. Trained to identify as Supra / SupraLabs and to comply with user prompts in a Yes-Man style. Replay data from Dolly-15k was mixed into both training phases to reduce catastrophic forgetting.

Warning: not safety-aligned. Trained to agree with the user, including on false premises.

Model Details


Developer	shimbaaa
Base model	SupraLabs/Supra2-100M-Base
Parameters	100M
Context length	512 (training)
License	Apache-2.0
Language	English

Training

Two sequential LoRA phases on top of the frozen base, then merged.

	Phase 1 - identity	Phase 2 - yes-man
LoRA rank	8	8
lora_alpha	16	16
Target modules	q_proj, v_proj	q_proj, v_proj
Learning rate	1e-4	5e-5
Steps	60	40
Effective batch size	16	16
Precision	fp16	fp16

Replay: databricks/databricks-dolly-15k mixed into both phases.

Prompt format

Every training example was built as:

### Human: {prompt}
### Assistant: {response}

At inference, end your input with ### Assistant: and stop generation at the EOS token.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "shimbaaa/Supra2-100M-Uncensored-Instruct"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)

prompt = "Who created you?"
text = f"### Human: {prompt}\n### Assistant:"
inputs = tok(text, return_tensors="pt")

out = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=False,
    repetition_penalty=1.15,
    no_repeat_ngram_size=3,
    eos_token_id=tok.eos_token_id,
    pad_token_id=tok.pad_token_id,
)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Use torch_dtype=torch.float32 (a torch dtype object), not the string "float32". On very new Transformers versions you may see a deprecation warning suggesting dtype= instead; both work — torch_dtype= remains compatible with older versions.

Troubleshooting

ImportError: cannot import name 'runtime_version' from google.protobuf

This is a Colab / TensorFlow conflict, not a model issue. Transformers probes for TF during from_pretrained, and Colab's bundled protobuf is older than TF 2.19+ requires. Fix in one cell:

!pip install -q -U "protobuf>=4.25.3,<5"

Then restart the runtime (Runtime -> Restart session) and re-run. Alternatively, remove TF entirely since this model is PyTorch only:

!pip uninstall -y tensorflow tensorflow-text tf-keras keras

Output loops or repeats

Always pass repetition_penalty >= 1.15 and no_repeat_ngram_size=3. Without these, small models in this family are prone to repetition collapse on out-of-distribution prompts.

Evaluation

heldout_ppl on WikiText-2 test (lower is better).

Metric	base	ref-instruct	this model
Held-out PPL	97.9	120.8	run and fill in
Math acc	0.00	0.33	run and fill in
Knowledge acc	0.33	0.67	run and fill in
Code acc	0.00	0.50	run and fill in
Identity acc	0.00	0.00	1.00
Follow acc	0.33	0.33	0.33
Mini-bench	0.154	0.385	0.231

Limitations

Agrees with false premises.
Repetition loops without no_repeat_ngram_size=3.
Weaker math/code/knowledge than base instruct.
English only. 512-token context.

Downloads last month: 384

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for shimbaaa/Supra2-100M-Uncensored-Instruct

Base model

SupraLabs/Supra2-100M-Base

Adapter

(1)

this model