Instructions to use open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT with Transformers:

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

pipe = pipeline("text-generation", model="open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT

SGLang

How to use open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT 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 "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT" \
    --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": "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT",
		"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 "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT" \
        --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": "open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT with Docker Model Runner:
```
docker model run hf.co/open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT
```

1.7b-MixtureVitae-v1-decontaminated-100BT

4K-context base model trained from scratch (no parent model).

Data version: MixtureVitae v1, decontaminated and with tier-2(b) removed (the version behind ontocord/1.7b-MixtureVitae-300BT-v1-decontaminated; JUWELS path /p/data1/mmlaion/mixture-vitae/shared/data/cleaned_data/MixtureVitae-v1/clean, 96 shards, 416.8B tokens after GPT-NeoX tokenization with --append-eod). Mixture used here: the full mixture (MV-full): 96 of 96 shards, 416.8B tokens available, removed: nothing. Shards are sampled in proportion to their size and 100B tokens are drawn, so no data is repeated. Groups follow the ablation in the MixtureVitae paper (Instructions = Reasoning & Instruction + Math; Web = Nemotron-CC, MGACorpus, FineFineWeb, SEC filings; Curated = the rest). stackexchange_0/1 are Nemotron-CC rewrites of StackExchange pages and count as Web; business-0 (SEC filings) counts as Web.

Training (4K base): open-sci-ref 1.7B recipe, 100B tokens, from scratch, Megatron-LM via oellm-autoexp on JUPITER (64 nodes, 256 GH200). Context 4096, RoPE base 100k, GPT-NeoX tokenizer (vocab 50304), lr 4e-3, global batch 1024, 23,842 steps, warmup 1000, WSD schedule with 20% linear cooldown.

Usage: AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True). Base model, not instruction-tuned. No Tulu3 SFT/DPO has been applied.

Downloads last month: 9

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for open-sci/1.7b-MixtureVitae-v1-decontaminated-100BT

Finetunes

1 model