OmicAI-4B

OmicAI is a family of language models optimized for omics data analysis and tool interaction — writing and running real bioinformatics pipelines over the omicverse toolkit. This card is the 4B member; 9B and 27B variants are in progress.

OmicAI is built on Qwen3.5 and fine-tuned with ms-swift using LoRA adapters merged into the final weights.

The OmicAI family

Model	Base	Access
OmicAI-0.8B	Qwen3.5-0.8B	Open — Apache-2.0
OmicAI-4B (this model)	Qwen3.5-4B	Open — Apache-2.0
OmicAI-9B	MiMo-V2.6-Distill-Qwen-9B	Open — Apache-2.0
OmicAI-27B	Qwen3.8-27B	By application

Training Data

The training trajectories are produced by OmicOS and OmicVerse through a paper-reproduction process over the last three years of publications: the systems retrieve recent omics papers, then attempt to reproduce their analyses end-to-end from the raw data. Each attempt yields a multi-turn agent trajectory — literature/context retrieval, code generation, omicverse tool calls, and recovery from execution feedback.

A trajectory is kept only when an automatic judge confirms the deliverable is legitimate (not a placeholder or answer-leak) and a scoreboard confirms the reproduced result matches or surpasses the paper's reported (SOTA) numbers. Runs that were rejected, contaminated, or degenerate are discarded. This yields a high-signal corpus grounded in real, verifiable scientific reproduction.

Key Details

Base model	Qwen/Qwen3.5-4B
Method	ms-swift 4.5.3 · LoRA (target_modules=all-linear, rank 16, α 32) → merged
Precision	bfloat16 (merged weights)
Chat template	qwen3_5 — inference template must match training
Training data	agent trajectories from OmicOS/OmicVerse reproducing 3 years of omics papers; judge + scoreboard (MATCH / SURPASS-SOTA) validated
Epochs / seq len	3 epochs · max length 16384 · packing + padding-free

Evaluation

OmicAI is evaluated on two axes (results reported separately as the suite lands):

• Domain / tool-use (the training target): an omics-analysis benchmark with a deterministic execution verifier, plus function-calling (BFCL) and multi-turn agentic (τ-bench) suites.
• General-capability regression (catastrophic forgetting check): MMLU, GSM8K, IFEval — to ensure domain gains do not come at the cost of general ability.

Benchmark numbers are omitted here rather than estimated; they will be added once the evaluation harness is finalized.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("omicverse/OmicAI-4B")
model = AutoModelForCausalLM.from_pretrained(
    "omicverse/OmicAI-4B", torch_dtype="bfloat16", device_map="auto")

msgs = [
    {"role": "system", "content": "You are an omics bioinformatics assistant."},
    {"role": "user", "content": "Load pbmc3k, run QC, normalize, and cluster with omicverse."},
]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))

Also loadable with vLLM (use the qwen3_5 chat template).

Intended Use & Limitations

• Intended: research assistant for single-cell / bulk / spatial omics analysis and omicverse pipeline generation.
• Limitations: a research model — outputs (especially generated analysis code) should be executed and verified before use; it does not replace expert judgement, and may hallucinate APIs outside the omicverse/scanpy surface it was trained on.

Ecosystem

Part of the omicOS / omicAI effort and the omicverse ecosystem. OmicAI is also the student model of omicOS-RSI, an infrastructure-level recursive self-improvement pipeline in which the model, the omicverse toolkit, and the task benchmark co-evolve under an execution-grounded verifier.

Citation

If omicverse is useful in your work, please cite the omicverse paper. A dedicated OmicAI report is forthcoming.