BioGravity-bilingual-Inst

BioGravity-bilingual-Inst is a Korean–English biomedical instruction model from AIGEN Sciences, Inc., developed from the Gravity 30B-A5B family. It is intended for biomedical explanations, instruction following, and tool-assisted research in the Biomni A1 environment.

The repository provides a complete merged checkpoint, tokenizer, chat template, Gravity model-loading code, and A1 protocol metadata.

Model Summary

Property	Value
Developer	AIGEN Sciences, Inc.
Base model	Gravity-30B-A5B-Preview
Architecture	GravityMoEForCausalLM (gravity_moe), sparse MoE with MLA
Total / active parameters	29.56B / approximately 5.34B
Layers / routed experts	52 / 64, with 8 experts selected per token
Primary languages	Korean and English
Configured context limit	131,072 tokens
Stored precision	FP32, 24 safetensors shards
Supported inference precision	BF16
Agent environment	Biomni A1
A1 protocol	Dedicated execute, solution, and observation special tokens
License	Apache 2.0

Evaluation Results

Reported Biomni Eval1 results from a separate evaluation environment:

Model	Evaluation harness	Accuracy
BioGravity-bilingual-Inst	Not applied	39.72%
BioGravity-bilingual-Inst	Applied	55.1%

The 55.1% result uses the evaluation harness. The harness and its configuration are available upon request through AIGEN Sciences.

Quickstart

Use a CUDA-compatible PyTorch installation and the tested Transformers version:

pip install "transformers==4.57.6" accelerate safetensors

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "aigensciences/BioGravity-bilingual-Inst"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="eager",
).eval()

messages = [
    {"role": "system", "content": "You are a biomedical research assistant."},
    {"role": "user", "content": "관찰연구에서 상관관계와 인과관계의 차이를 한국어로 설명해 주세요."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    tokenizer_kwargs={"return_token_type_ids": False},
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=1024, do_sample=False)

continuation = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(continuation, skip_special_tokens=False))

BF16 loading requires approximately 59 GB for weights, plus memory for the KV cache and runtime. The configured 131,072-token context limit does not establish benchmark performance at that length.

Biomni A1 integration

Prepare the tools and data lake using the Biomni installation instructions, together with the evaluation harness supplied on request.

Purpose	Format
Reasoning segment	<think>...</think>
Code action	<execute>Python code</execute>
Tool observation	<observation>tool result</observation>
Final answer	<solution>answer</solution>

The A1 runtime executes actions and returns observations. Preserve the six dedicated action/observation special tokens by using skip_special_tokens=False. The runtime should retain the </execute> and </solution> stop markers when delimiting actions and final answers.

The default generation prompt opens a <think> segment. Passing enable_thinking=False to apply_chat_template prefills an empty <think></think> segment. Keep the tokenizer, chat template, and protocol_manifest.json together when deploying the model.

Limitations

• Research use and environment. The model is released for research purposes and is intended to operate in the Biomni A1 environment.
• Factual accuracy. Generated content, explanations, and references may contain factual errors. Expert verification is required before relying on the results.
• Knowledge freshness. Literature and guidelines published after the training-data cutoff are not reflected in the model weights. Current sources need to be retrieved and verified separately.
• Language coverage. Performance in languages other than Korean and English is not guaranteed.
• Harness dependence. The reported 55.1% Eval1 result uses the evaluation harness and environment associated with that measurement. Performance depends on the deployment and evaluation setup.

Acknowledgements

Support for this research comes from the 인공지능 특화 파운데이션 모델 프로젝트 (Domain-Specific Foundation Model Project). 과학기술정보통신부 (MSIT) provides funding, and 정보통신산업진흥원 (NIPA) administers the program.

We thank Lunit, Trillion Labs, and the consortium's industry, academic, and hospital partners for the Gravity research collaboration. The Lunit model card provides the consortium acknowledgements on which this statement is based. We also acknowledge the Biomni project for its agent framework and evaluation resources.

License

The model is released under the Apache License 2.0. Research use describes its intended application and does not add restrictions to the license.

Contact

AIGEN Sciences, Inc. — contact the organization to request the evaluation harness and configuration.