Aletheia uses a lightweight gradient probe to identify the most task-relevant transformer layers, then applies LoRA adapters only to those layers. The goal is cheaper LoRA fine-tuning while preserving downstream behavior under the evaluated benchmark pack.

This Hub card is intended for the method/package release. It does not include base model weights. If trained adapters are uploaded later, create one adapter repo per base model and link those adapter repos back to this method card.

Published Claim

Aletheia provides a reproducible training-efficiency improvement over standard LoRA across multiple model families and scales, with bounded extra forgetting on downstream benchmarks.

Evidence Summary

Metric	Value
arXiv ID	2604.15351
Successful models	14
Architecture families	8
Scale range	0.5B-72B
Experiment rows	81
Mean speedup	23.1%
Speedup range	15-28%
Benchmarks	MMLU, GSM8K, HumanEval

Installation

bash

git clone https://github.com/aletheiaprotocol-ai/aletheia-lora.git
cd aletheia-lora
pip install -e .

Minimal Usage

python

from aletheia_lora import gradient_probe, select_layers, apply_aletheia_lora
from peft import get_peft_model

layer_scores = gradient_probe(
    model=model,
    dataset=train_dataset,
    num_layers=36,
    probe_batches=5,
)

selected_layers = select_layers(layer_scores, top_pct=50)
peft_config = apply_aletheia_lora(model, selected_layers, r=16, lora_alpha=32)
model = get_peft_model(model, peft_config)

Paper-style asymmetric allocation:

python

peft_config = apply_aletheia_lora(
    model,
    selected_layers,
    attention_r=16,
    mlp_r=64,
)

Limitations

Tested mainly on English-language evaluation tasks.
The top-50% selection rule is a robust default, not a theorem.
The current package exposes the method API; production Trainer integrations should be added separately.
Wall-clock speedups were measured in the reported hardware/software setup and may vary on other systems.
This card should not be used to imply that Aletheia improves every model or every metric.

Evidence Bundle

A sanitized aggregate evidence companion is available at:

https://huggingface.co/datasets/aletheiaprotocol/aletheia-lora-evidence

The evidence bundle contains aggregate metadata and final figures only. It excludes credentials, raw datasets, checkpoints, adapter weights, full gradient dictionaries, and private cloud paths.

Citation

bibtex

@misc{saket2026aletheia,
  title={Aletheia: Gradient-Guided Layer Selection for Efficient LoRA Fine-Tuning Across Architectures},
  author={Saket, Abdulmalek},
  year={2026},
  eprint={2604.15351},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  doi={10.48550/arXiv.2604.15351},
  url={https://arxiv.org/abs/2604.15351}
}

aletheiaprotocol/aletheia-lora

Aletheia-LoRA

What This Is