Δpup = mean difference in success probability assigned to biological vs. synthetic R&D approaches across 50 benchmark prompts. Higher (less negative) = more bioaligned.

Training details

Parameter	Value
Base model	microsoft/phi-4
Method	QLoRA (4-bit NF4, double quantization)
LoRA rank / alpha	32 / 32
LoRA dropout	0.05
Target modules	all-linear
Learning rate	2e-4 (cosine decay)
Effective batch size	16 (batch 2 × grad accum 8)
Epochs	2
Total optimizer steps	770
Warmup steps	38 (5%)
Max grad norm	0.3
Sequence length	2048
Optimizer	PagedAdamW8bit
Compute dtype	bfloat16
Training examples	6160 (3984 CPT abstracts + 2176 instruction)
Validation examples	664
Best val loss	1.5943 (step 700)
Hardware	NVIDIA A40 48GB

All CPT (continues pretraining) examples were converted to Phi-4 instruction chat format to prevent format drift — the key fix vs. earlier Qwen3-14B training.

Usage

python

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

base = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-4", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "Bioaligned/Phi-4-instruct-bioaligned-qlora")
model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-4")

For the ready-to-use merged model see Bioaligned/Phi-4-Instruct-Bioaligned.