Instructions to use BEE-spoke-data/verysmol_llama-v11-KIx2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BEE-spoke-data/verysmol_llama-v11-KIx2 with Transformers:

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

pipe = pipeline("text-generation", model="BEE-spoke-data/verysmol_llama-v11-KIx2")

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("BEE-spoke-data/verysmol_llama-v11-KIx2")
model = AutoModelForMultimodalLM.from_pretrained("BEE-spoke-data/verysmol_llama-v11-KIx2")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use BEE-spoke-data/verysmol_llama-v11-KIx2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BEE-spoke-data/verysmol_llama-v11-KIx2"
# 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/verysmol_llama-v11-KIx2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/BEE-spoke-data/verysmol_llama-v11-KIx2

SGLang

How to use BEE-spoke-data/verysmol_llama-v11-KIx2 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/verysmol_llama-v11-KIx2" \
    --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/verysmol_llama-v11-KIx2",
		"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/verysmol_llama-v11-KIx2" \
        --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/verysmol_llama-v11-KIx2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use BEE-spoke-data/verysmol_llama-v11-KIx2 with Docker Model Runner:
```
docker model run hf.co/BEE-spoke-data/verysmol_llama-v11-KIx2
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

verysmol_llama-v11-KIx2

Model description

This model is a fine-tuned version of v10 (refinedweb-3m dedup) further trained for 2 epochs on KI dataset.

It achieves the following results on the evaluation set:

Loss: 2.8876
Accuracy: 0.4502

evals

hf-causal-experimental (pretrained=pszemraj/verysmol_llama-v11-KIx2,revision=main,trust_remote_code=True,dtype='float'), limit: None, provide_description: False, num_fewshot: 0, batch_size: 16

Task	Version	Metric	Value		Stderr
arc_easy	0	acc	0.4024	±	0.0101
		acc_norm	0.3788	±	0.0100
boolq	1	acc	0.6199	±	0.0085
lambada_openai	0	ppl	111.9939	±	4.6906
		acc	0.2354	±	0.0059
openbookqa	0	acc	0.1440	±	0.0157
		acc_norm	0.2760	±	0.0200
piqa	0	acc	0.5713	±	0.0115
		acc_norm	0.5664	±	0.0116
winogrande	0	acc	0.5201	±	0.0140

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	0.1971	±	0.0116
		acc_norm	0.2278	±	0.0123

Task	Version	Metric	Value		Stderr
hellaswag	0	acc	0.2618	±	0.0088
		acc_norm	0.2797	±	0.0090

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	0.2509	±	0.0152
		mc2	0.4492	±	0.0156

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.00014
train_batch_size: 16
eval_batch_size: 16
seed: 17514
gradient_accumulation_steps: 8
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-06
lr_scheduler_type: inverse_sqrt
lr_scheduler_warmup_ratio: 0.05
num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
3.0681	0.03	150	3.0689	0.4259
3.0113	0.07	300	3.0433	0.4278
2.9468	0.1	450	3.0362	0.4288
3.0162	0.13	600	3.0148	0.4326
2.9531	0.17	750	3.0012	0.4341
2.9282	0.2	900	2.9923	0.4358
2.9485	0.23	1050	2.9845	0.4357
2.9365	0.27	1200	2.9749	0.4375

...

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.8215	1.7	7650	2.8943	0.4496
2.7714	1.74	7800	2.8914	0.4501
2.8132	1.77	7950	2.8913	0.4500
2.8505	1.8	8100	2.8906	0.4502
2.8294	1.84	8250	2.8901	0.4502
2.7977	1.87	8400	2.8891	0.4499
2.7501	1.9	8550	2.8878	0.4505
2.8038	1.94	8700	2.8883	0.4504
2.7547	1.97	8850	2.8876	0.4502

Downloads last month: 66

Safetensors

Model size

58.1M params

Tensor type

F32

Model tree for BEE-spoke-data/verysmol_llama-v11-KIx2

Quantizations

2 models

Collection including BEE-spoke-data/verysmol_llama-v11-KIx2

smol llama

Collection

🚧"raw" pretrained smol_llama checkpoints - WIP 🚧 • 4 items • Updated Apr 29, 2024 • 6