omerfarooq223

gemma-4-12b-evidence-verification-qlora

Deploy Dedicated

README

License: apache-2.0

Model Details

Model Description

Developed by: Muhammad Umar Farooq
Shared by: Muhammad Umar Farooq
Model type: Fine-tuned PEFT/QLoRA Parameter-Efficient Adapter
Language(s) (NLP): English (en)
License: Apache 2.0
Finetuned from model: google/gemma-4-12B-it

Model Sources

Repository: https://github.com/omerfarooq223/gemma-claim-verification
Demo Space: https://huggingface.co/spaces/omerfarooq223/gemma-claim-verifier

Uses

Direct Use

Given a natural language claim and one or more supplied evidence passages, classify the evidence relation into exactly one of three categories:

SUPPORTS
REFUTES
NOT_ENOUGH_INFO

Downstream Use

Fact-checking applications, automated RAG verification systems, hallucination detection engines, and evidence audit tools.

Out-of-Scope Use

General factual QA without evidence passages. The model is calibrated strictly to evaluate truth conditions only relative to the supplied text passages.

Risk, Bias, and Limitations

Bounded Context: Prompts and evidence passages must fit within 256 tokens.
Evidence-Only Scope: The model does not query external knowledge bases or search the web; it evaluates factual claims strictly grounded in the supplied context.
Error Pattern Analysis: Achieved 99.30% precision on REFUTES and 100% recall on NOT_ENOUGH_INFO. Subtle numerical or scope contradictions with supportive phrasing can occasionally be mispredicted as SUPPORTS.

How to Get Started with the Model

python

import torchfrom transformers import AutoProcessor, Gemma4UnifiedForConditionalGeneration, BitsAndBytesConfigfrom peft import PeftModel # 1. Load 4-bit NF4 quantized base modelbnb_config = BitsAndBytesConfig(    load_in_4bit=True,    bnb_4bit_quant_type="nf4",    bnb_4bit_use_double_quant=True,    bnb_4bit_compute_dtype=torch.float16,) processor = AutoProcessor.from_pretrained("google/gemma-4-12B-it")base_model = Gemma4UnifiedForConditionalGeneration.from_pretrained(    "google/gemma-4-12B-it",    quantization_config=bnb_config,    device_map="auto",    torch_dtype=torch.float16,) # 2. Attach Winning LoRA Adaptermodel = PeftModel.from_pretrained(    base_model,    "omerfarooq223/gemma-4-12b-evidence-verification-qlora",)

Training Details

Training Data

Audited Clean Base: 935 semantically audited examples (cleaned from 1,000 raw noisy rows via a 10-step audit pipeline).
Audited Contrastive Curriculum: 150 contrastive examples arranged in trios.
Total Training Dataset: 1,085 curated examples.

Training Hyperparameters

Quantization: 4-bit NF4 (bnb_4bit_use_double_quant=True)
Compute Precision: FP16 (torch.float16)
LoRA Config: Rank $r = 8$ , Alpha $α = 16$ , Dropout $0.05$
Target Modules: q_proj, k_proj, v_proj, o_proj, , ,

Evaluation Benchmark Performance

Supervised Event-Day Test Set (500 Examples)

$Overall Accuracy : 94.40% (472/500 correct) Macro-F1 : 94.38%$

Table with columns: Label Class, Precision, Recall, F1-Score, Support (Ground Truth)
Label Class	Precision	Recall	F1-Score	Support (Ground Truth)
`SUPPORTS`	86.39%	98.80%	92.18%	167
`REFUTES`	99.30%	84.43%	91.26%	167
`NOT_ENOUGH_INFO`

Citation

bibtex

@misc{farooq2026gemma4claimverification,  author = {Farooq, Muhammad Umar},  title = {Reliable Evidence-Based Claim Verification with Gemma 4 12B},  year = {2026},  publisher = {Hugging Face},  howpublished = {\url{https://huggingface.co/omerfarooq223/gemma-4-12b-evidence-verification-qlora}}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

omerfarooq223

Model Tree

Base

google/gemma-4-12B-it

Adapter

this model

Input Modalities

TextAudioImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Model Description

Developed by: Muhammad Umar Farooq
Shared by: Muhammad Umar Farooq
Model type: Fine-tuned PEFT/QLoRA Parameter-Efficient Adapter
Language(s) (NLP): English (en)
License: Apache 2.0
Finetuned from model: google/gemma-4-12B-it

Model Sources

Repository: https://github.com/omerfarooq223/gemma-claim-verification
Demo Space: https://huggingface.co/spaces/omerfarooq223/gemma-claim-verifier

Uses

Direct Use

Given a natural language claim and one or more supplied evidence passages, classify the evidence relation into exactly one of three categories:

SUPPORTS
REFUTES
NOT_ENOUGH_INFO

Downstream Use

Fact-checking applications, automated RAG verification systems, hallucination detection engines, and evidence audit tools.

Out-of-Scope Use

General factual QA without evidence passages. The model is calibrated strictly to evaluate truth conditions only relative to the supplied text passages.

Risk, Bias, and Limitations

Bounded Context: Prompts and evidence passages must fit within 256 tokens.
Evidence-Only Scope: The model does not query external knowledge bases or search the web; it evaluates factual claims strictly grounded in the supplied context.
Error Pattern Analysis: Achieved 99.30% precision on REFUTES and 100% recall on NOT_ENOUGH_INFO. Subtle numerical or scope contradictions with supportive phrasing can occasionally be mispredicted as SUPPORTS.

How to Get Started with the Model

python

import torchfrom transformers import AutoProcessor, Gemma4UnifiedForConditionalGeneration, BitsAndBytesConfigfrom peft import PeftModel # 1. Load 4-bit NF4 quantized base modelbnb_config = BitsAndBytesConfig(    load_in_4bit=True,    bnb_4bit_quant_type="nf4",    bnb_4bit_use_double_quant=True,    bnb_4bit_compute_dtype=torch.float16,) processor = AutoProcessor.from_pretrained("google/gemma-4-12B-it")base_model = Gemma4UnifiedForConditionalGeneration.from_pretrained(    "google/gemma-4-12B-it",    quantization_config=bnb_config,    device_map="auto",    torch_dtype=torch.float16,) # 2. Attach Winning LoRA Adaptermodel = PeftModel.from_pretrained(    base_model,    "omerfarooq223/gemma-4-12b-evidence-verification-qlora",)

Training Details

Training Data

Audited Clean Base: 935 semantically audited examples (cleaned from 1,000 raw noisy rows via a 10-step audit pipeline).
Audited Contrastive Curriculum: 150 contrastive examples arranged in trios.
Total Training Dataset: 1,085 curated examples.

Training Hyperparameters

Quantization: 4-bit NF4 (bnb_4bit_use_double_quant=True)
Compute Precision: FP16 (torch.float16)
LoRA Config: Rank $r = 8$ , Alpha $α = 16$ , Dropout $0.05$
Target Modules: q_proj, k_proj, v_proj, o_proj, , ,

Evaluation Benchmark Performance

Supervised Event-Day Test Set (500 Examples)

$Overall Accuracy : 94.40% (472/500 correct) Macro-F1 : 94.38%$

Table with columns: Label Class, Precision, Recall, F1-Score, Support (Ground Truth)
Label Class	Precision	Recall	F1-Score	Support (Ground Truth)
`SUPPORTS`	86.39%	98.80%	92.18%	167
`REFUTES`	99.30%	84.43%	91.26%	167
`NOT_ENOUGH_INFO`

Citation

bibtex

@misc{farooq2026gemma4claimverification,  author = {Farooq, Muhammad Umar},  title = {Reliable Evidence-Based Claim Verification with Gemma 4 12B},  year = {2026},  publisher = {Hugging Face},  howpublished = {\url{https://huggingface.co/omerfarooq223/gemma-4-12b-evidence-verification-qlora}}}

@misc{farooq2026gemma4claimverification,  author = {Farooq, Muhammad Umar},  title = {Reliable Evidence-Based Claim Verification with Gemma 4 12B},  year = {2026},  publisher = {Hugging Face},  howpublished = {\url{https://huggingface.co/omerfarooq223/gemma-4-12b-evidence-verification-qlora}}}