bahnator-1.1-pro-preview

LoRA adapters for LiquidAI/d1-3B that check an agent's decision card for known traps before a person sees it: a stale GPS fix, a replaced instruction, a wrong escalation. They read text and photos.

Built for Controlane, a control room for car delivery, at the Prompt Operations Hackathon (24 to 25 Oct 2026). Live at https://controlane.bobs.build.

Results

On 4,395 synthetic test cases from 203 template families held out of training completely:

Checker	Catch rate	False alarms	Time per check
bahnator (this model)	92%	5%	22 ms
Jev (typesafe/jev-1.14)	92%	13%	229 ms
Laya v6, fine-tuned	86%	8%	10 ms
d1-3B, as released	67%	33%	23 ms

93.5% correct overall. 32 of 34 on the hand-labelled set, 216 of 220 on photos and scans.

Training

• LoRA rank 16, alpha 32, on q_proj k_proj v_proj out_proj in_proj w1 w2 w3; 24.5 M trainable parameters
• 3 epochs, AdamW, learning rate 2e-4
• Synthetic train split only; no test case was used
• 24 minutes on one RTX 6000 Ada on Verda (Finland), about $1.33
• Best of 7 training runs across 2 models

Use

d1-lora-v1.pt holds rank, alpha, targets and the A/B weights per layer. train_d1.py has add_lora and merge_lora:

import torch
from transformers import AutoModel
from train_d1 import add_lora, merge_lora

model = AutoModel.from_pretrained("LiquidAI/d1-3B", trust_remote_code=True, dtype=torch.bfloat16).to("cuda")
saved = torch.load("d1-lora-v1.pt")
lm = model.model.language_model
for name, lora in add_lora(lm, saved["rank"], saved["alpha"]).items():
    lora.A.data.copy_(saved["weights"][name]["A"])
    lora.B.data.copy_(saved["weights"][name]["B"])
merge_lora(lm)
model.eval()
model.system_one(state, questions, images=images)

All training data is synthetic and fictional.