baogiadoan/Llama-3.1-8B-Instruct-DoRA-adapter

Llama-3.1-8B-Instruct — DoRA adapter

Built with Llama.

A LoRA adapter for meta-llama/Llama-3.1-8B-Instruct, fine-tuned on the synthetic training split of DoRA (Domain-oriented RAG Assessment) for grounded question answering over specialist-domain documents.

📄 Paper · 💻 Code · 📊 Dataset

Naming note. "DoRA" here is the framework — Domain-oriented RAG Assessment. This adapter uses standard LoRA, not Weight-Decomposed Low-Rank Adaptation (Liu et al., 2024).

Results

Evaluated on the 1,259-instance DoRA test set with GTE retrieval (top-k=3). Mean ± std over three fine-tuning seeds; baselines over three inference runs.

Model	ROUGE-L-R ↑	Tok-R ↑	Completeness ↑	Hallucination ↓
Llama-3.1-8B-Instruct (base)	63.30±0.80	62.76±0.74	66.72±0.51	4.75±0.23
This adapter	69.12±0.19	71.13±0.16	73.37±0.40	2.17±0.35
Gain vs base	+5.82	+8.37	+6.65	+2.58

Every gain is significant under a paired bootstrap test (10,000 resamples, p < 0.001). Hallucination is more than halved. Gains persist across alternative retrievers, gold evidence, and matched answer-length budgets.

Training

Base model	meta-llama/Llama-3.1-8B-Instruct
Method	LoRA (PEFT)
Rank / alpha / dropout	64 / 128 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj
Epochs	3 (checkpoint 900)
Learning rate	1e-5
Batch size	4 × grad-accum 4
Max sequence length	4096
Precision	bf16, gradient checkpointing
Training data	5,052 train / 266 validation, generated by Claude Sonnet over a seed corpus disjoint from the test corpus

Three seeds are released. Seed 42 (the primary model reported in the paper) is at the repository root; seeds 2 and 3 are under seed2/ and seed3/.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "meta-llama/Llama-3.1-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "baogiadoan/Llama-3.1-8B-Instruct-DoRA-adapter")
tokenizer = AutoTokenizer.from_pretrained(base_id)

The model expects the grounded prompt used throughout the paper — retrieved context, then the question:

You are provided with grounded evidence. Use only this context to answer.
If the evidence is insufficient, respond with "I don't have enough context."

Context:
{context}

Question:
{question}

Answer:

That template is released as prompts/evaluation/grounded_inference.yaml in the code repository. Behaviour will differ if you prompt it another way.

Intended use and limitations

Built for grounded, evidence-attributed QA over specialist-domain documents. It is trained to answer from supplied context and to decline when the evidence is insufficient — it is not a general-purpose assistant and should not be used as a knowledge source without retrieval.

The training corpus is Australian defence-related public policy documents. The paper's evaluation is in that domain, and transfer to other specialist domains is untested. Answers are markedly more verbose than the reference answers (length ratio ≈ 5×); the paper includes a truncation analysis showing the coverage gains survive matched length budgets.

This is a research artifact. It has not been safety-tuned beyond the base model's alignment, and it should not be used for operational decision-making.

Licence

Governed by the Llama 3.1 Community License as a derivative of meta-llama/Llama-3.1-8B-Instruct. You must comply with that agreement and the Llama 3.1 Acceptable Use Policy. Built with Llama.

Citation

@misc{doan2026benchmarkconstructionevaluationframework,
      title={A Benchmark Construction and Evaluation Framework for Specialist Domains: Case Study on Defense-related Documents},
      author={Bao Gia Doan and Aditya Joshi and Pantelis Elinas and Aarya Bodhankar and Oscar Leslie and Tom Marchant and Flora Salim},
      year={2026},
      eprint={2604.17943},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.17943},
}