Instructions to use open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096 with Transformers:

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

pipe = pipeline("text-generation", model="open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096

SGLang

How to use open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096 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/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096" \
    --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": "open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096",
		"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 "open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096" \
        --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": "open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096 with Docker Model Runner:
```
docker model run hf.co/open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096
```

open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-ftrest-300B-4096

A 1.7B open-sci-ref model pretrained for 300B tokens at 4,096 tokens of context on MixtureVitae v1 (decontaminated, tier2b removed) without the instruction and reasoning subset ("noinstruct"), with the formatted-text shards (formatted_text-0 and formatted_text-1) added back ("ftrest"). It is a base model: no long-context extension, no supervised or preference fine-tuning. It sits beside open-sci/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-300B-4096, the same recipe without the formatted-text shards.

Training

Trained by Harsh Raj on JUPITER (64 nodes, 256 GH200) with oellm-autoexp and Megatron-LM, as the run opensci-ref_1.7b_MixtureVitae-v1-decontaminated-noinstruct-ftrest_300B_ctx-4096_rope-100k. This repository holds its final checkpoint, iteration 71,526.

Setting	Value
Data	MixtureVitae v1 decontaminated, tier2b removed, without the instruction and reasoning subset, with `formatted_text-0/1`: 58 shards, 242.1B tokens
Tokens	300B (71,526 iterations × 1,024 sequences × 4,096 tokens)
Architecture	24 layers, hidden size 2,048, FFN 8,192, 32 attention heads with 6 key-value groups, RMSNorm, QK layer norm, RoPE with theta 100,000
Context	4,096 tokens
Optimizer	AdamW (β2 0.95), peak learning rate 0.004, WSD schedule: 25,000 warmup iterations, linear decay over the last 14,305
Precision	bf16
Tokenizer	GPT-NeoX

The weights were converted to Hugging Face format by the run's own postprocessing (hf/iter_0071526); model.safetensors is that file unchanged (SHA-256 928314ffebd6b67190f0a38b029f98bf07d733be9ec6031a6dc02874830df7fa).

Files

Two files were added to the conversion so that the repository matches the copy the evaluations used, as Ali Elganzory's evaluation copies of the other 4K bases (ali-elganzory/…) do; the open-sci/ noinstruct repository above has the converter's files without them:

chat_template.jinja: the base-model chat template of ali-elganzory/open-sci-ref-v0.02-1.7b-mixturevitae-noinstruct-300B-4096, which the Protocol A reasoning evaluations apply.
modeling_opensci.py: the version from the same repository. It differs from the converter's only in passing **kwargs through to attention.

Load with trust_remote_code=True.

Evaluation

Evaluated with the project's Protocol A reasoning suite and the open-sci-0.01 language suite, three seeds each; the setup, calibration and results are in docs/investigations/protocol_a_jureca.md of the MixtureVitae post-training repository.

Downloads last month: 355

Safetensors

Model size

2B params

Tensor type

BF16