lokahq/Trinity-Mini-AI-Scientist

Trinity-Mini-AI-Scientist

Trinity-Mini-AI-Scientist is a LoRA adapter for arcee-ai/Trinity-Mini, post-trained with Group Relative Policy Optimization (GRPO) for scientific tool use, drug-discovery workflows, and Gene Ontology reasoning.

This is an adapter, not a standalone model; the base Trinity Mini weights are required.

Intended use

The adapter is designed for research workflows that combine:

• grounded use of drug-discovery and biomedical tools;
• evidence-backed synthesis from tool results;
• protein-function reasoning and structured Gene Ontology prediction;
• multi-agent AI-scientist applications.

It is a research model and must not be used as a substitute for qualified medical, clinical, or laboratory judgment.

Training

The adapter was trained for 100 GRPO steps with a 16,384-token training sequence length and a balanced mixture of two environments:

• lokahq/drug-tool-rl@1
• lokahq/bioreason-go-rl@1

Setting	Value
LoRA rank / alpha	64 / 128
Batch size / group size	256 / 16
Optimizer	Muon
Learning rate	3e-5
Scheduler	Cosine, 8 warmup steps, 5e-6 minimum LR
KL coefficient	1e-3
Maximum tool turns	12 (training), 8 (evaluation)

Evaluation

Accuracy is reported as Avg@1: the score from a single model response on each held-out evaluation example.

Evaluation	Metric	Result
Drug Tool	Avg@1	0.812
BioReason	Avg@1	0.863

Usage

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "arcee-ai/Trinity-Mini"
adapter_id = "lokahq/Trinity-Mini-AI-Scientist"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, adapter_id)