Instructions to use BEE-spoke-data/mega-ar-126m-4k with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BEE-spoke-data/mega-ar-126m-4k with Transformers:

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

pipe = pipeline("text-generation", model="BEE-spoke-data/mega-ar-126m-4k")

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("BEE-spoke-data/mega-ar-126m-4k", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use BEE-spoke-data/mega-ar-126m-4k with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BEE-spoke-data/mega-ar-126m-4k"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BEE-spoke-data/mega-ar-126m-4k",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/BEE-spoke-data/mega-ar-126m-4k

SGLang

How to use BEE-spoke-data/mega-ar-126m-4k 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 "BEE-spoke-data/mega-ar-126m-4k" \
    --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": "BEE-spoke-data/mega-ar-126m-4k",
		"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 "BEE-spoke-data/mega-ar-126m-4k" \
        --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": "BEE-spoke-data/mega-ar-126m-4k",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use BEE-spoke-data/mega-ar-126m-4k with Docker Model Runner:
```
docker model run hf.co/BEE-spoke-data/mega-ar-126m-4k
```

BEE-spoke-data/mega-ar-126m-4k

This may not be the best language model, but it is a language model! It's interesting for several reasons, not the least of which is that it's not technically a transformer.

Details:

768 hidden size, 12 layers
no MEGA chunking, 4096 context length
EMA dimension 16, shared dimension 192
tokenizer: GPT NeoX
train-from-scratch

For more info on MEGA (& what some of the params above mean), check out the model docs or the original paper

Open BEE-spoke-data/mega-ar-126m-4k in hfviewer

A more detailed and useful view (based on expanding the viewer + some reformatting)

Usage

Usage is the same as any other small textgen model.

Given the model's small size and architecture, it's probably best to leverage its longer context by adding input context to "see more" rather than "generate more".

evals

Initial data:

hf-causal-experimental (pretrained=BEE-spoke-data/mega-ar-126m-4k,revision=main,trust_remote_code=True,dtype='float'), limit: None, provide_description: False, num_fewshot: 0, batch_size: 4

Task	Version	Metric	Value		Stderr
arc_easy	0	acc	0.4415	±	0.0102
		acc_norm	0.3969	±	0.0100
boolq	1	acc	0.5749	±	0.0086
lambada_openai	0	ppl	94.9912	±	3.9682
		acc	0.2408	±	0.0060
openbookqa	0	acc	0.1660	±	0.0167
		acc_norm	0.2780	±	0.0201
piqa	0	acc	0.5974	±	0.0114
		acc_norm	0.5914	±	0.0115
winogrande	0	acc	0.4830	±	0.0140

Downloads last month: 441

Safetensors

Model size

0.1B params

Tensor type

F32

Datasets used to train BEE-spoke-data/mega-ar-126m-4k

Paper for BEE-spoke-data/mega-ar-126m-4k

Mega: Moving Average Equipped Gated Attention

Paper • 2209.10655 • Published Sep 21, 2022 • 1