README.md

---
base_model: google/gemma-3-4b-it
library_name: peft
pipeline_tag: text-generation
language:
- en
tags:
- peft
- lora
- data-to-text
- text-to-data
- factual-consistency
- hallucination-detection
---

# XQDT E2E verifier: gemma3 4B

This repository contains the LoRA adapter for the **gemma3 4B**
XQDT verifier from *XQDT: eXplainable and Quantitative Data-Text Alignment Metric
with Feedback Signals*. It is used with
[`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it).

This E2E checkpoint was trained on the joint WebNLG--E2E synthetic training set.

## Overview

XQDT verifies alignment between English text and structured triples. It returns
`missing`, `extra`, and `incorrect` units, or `All correct`. `missing` identifies
an input unit omitted from the text, `extra` identifies text content unsupported
by the input, and `incorrect` identifies an input unit realised with incorrect
information.

Example inputs and outputs are provided in `smoke_test.json`.
Generated text may vary slightly across inference libraries and package versions.

## Prompt format

```text
Verify if the triples align with the text. Find missing, extra, or incorrect triples.
TEXT: {text}
TRIPLES:
1. [S] {subject} [P] {predicate} [O] {object}
Output as markdown table with Type and Triple columns.
```

## ms-swift

```python
import torch
from swift.infer_engine import InferRequest, RequestConfig, TransformersEngine

BASE_MODEL = "google/gemma-3-4b-it"
ADAPTER_ID = "Loria-MosAIk/xqdt-e2e-gemma3-4b"
SYSTEM_PROMPT = "Identify extra, missing and incorrect triples precisely."
QUERY = """Verify if the triples align with the text. Find missing, extra, or incorrect triples.
TEXT: Blue Spice is a coffee shop in city centre.
TRIPLES:
1. [S] Blue Spice [P] area [O] city centre
2. [S] Blue Spice [P] eat type [O] coffee shop
Output as markdown table with Type and Triple columns."""
MESSAGES = [{"role": "user", "content": QUERY}]

engine = TransformersEngine(
    BASE_MODEL,
    adapters=[ADAPTER_ID],
    max_batch_size=1,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    template_type="gemma3_text",
    use_hf=True,
)
response = engine.infer(
    [InferRequest(messages=MESSAGES)],
    RequestConfig(max_tokens=1024, temperature=0.3, seed=2023),
    use_tqdm=False,
)[0]
print(response.choices[0].message.content)
```

## Transformers and PEFT

```python
import torch
from peft import PeftModel
from transformers import set_seed

BASE_MODEL = "google/gemma-3-4b-it"
ADAPTER_ID = "Loria-MosAIk/xqdt-e2e-gemma3-4b"
SYSTEM_PROMPT = "Identify extra, missing and incorrect triples precisely."
QUERY = """Verify if the triples align with the text. Find missing, extra, or incorrect triples.
TEXT: Blue Spice is a coffee shop in city centre.
TRIPLES:
1. [S] Blue Spice [P] area [O] city centre
2. [S] Blue Spice [P] eat type [O] coffee shop
Output as markdown table with Type and Triple columns."""
MESSAGES = [{"role": "user", "content": QUERY}]
set_seed(2023)

from transformers import AutoProcessor, Gemma3ForConditionalGeneration

processor = AutoProcessor.from_pretrained(BASE_MODEL)
base = Gemma3ForConditionalGeneration.from_pretrained(
    BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, ADAPTER_ID).eval()
prompt = processor.apply_chat_template(MESSAGES, tokenize=False, add_generation_prompt=True)
inputs = processor(text=prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.3)
generated = output[0, inputs["input_ids"].shape[-1]:]
print(processor.decode(generated, skip_special_tokens=True))
```

## vLLM

```python
from huggingface_hub import snapshot_download
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

BASE_MODEL = "google/gemma-3-4b-it"
ADAPTER_ID = "Loria-MosAIk/xqdt-e2e-gemma3-4b"
SYSTEM_PROMPT = "Identify extra, missing and incorrect triples precisely."
QUERY = """Verify if the triples align with the text. Find missing, extra, or incorrect triples.
TEXT: Blue Spice is a coffee shop in city centre.
TRIPLES:
1. [S] Blue Spice [P] area [O] city centre
2. [S] Blue Spice [P] eat type [O] coffee shop
Output as markdown table with Type and Triple columns."""
MESSAGES = [{"role": "user", "content": QUERY}]

adapter_path = snapshot_download(ADAPTER_ID)
llm = LLM(model=BASE_MODEL, enable_lora=True)
outputs = llm.chat(
    MESSAGES,
    SamplingParams(max_tokens=1024, temperature=0.3, seed=2023),
    lora_request=LoRARequest("xqdt", 1, adapter_path),
)
print(outputs[0].outputs[0].text)
```

## Citation

```bibtex
@inproceedings{efimov-zhang-etal-2026-xqdt,
  title = {XQDT: eXplainable and Quantitative Data-Text Alignment Metric with Feedback Signals},
  author = {Efimov-Zhang, Kun and Song, Yifei and Gardent, Claire},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year = {2026}
}
```