Downloads · 30 days
11
2% of all-time downloads
ANMOLGPT-4B-v0.3
anmoldhandhania93/ANMOLGPT-4B-v0.3
ANMOLGPT-4B-v0.3 is a text generation model from anmoldhandhania93. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
ANMOLGPT-4B-v0.3 is the third release in the ANMOLGPT series. Built on Qwen3.5-4B, this version focuses on improving factual accuracy and reasoning through targeted fine-tuning on the TruthfulQA dataset.
All-time downloads
444
Public
Parameters
4.7B
9.3 GB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors9.3 GB · 100%
Parameter types
How the weights are stored.
BF164.7B · 100%
Try a prompt
Base models
Model card
From the Hugging Face model README
ANMOLGPT-4B-v0.3
ANMOLGPT-4B-v0.3 is the third release in the ANMOLGPT series. Built on Qwen3.5-4B, this version focuses on improving factual accuracy and reasoning through targeted fine-tuning on the TruthfulQA dataset.
Compared to previous versions, v0.3 demonstrates measurable improvements across multiple reasoning benchmarks while maintaining strong general-purpose capabilities.
Highlights
- Built on Qwen3.5-4B
- Fine-tuned using LoRA with Unsloth
- Trained on TruthfulQA
- Improved reasoning performance
- Improved factual accuracy
- Open-source and Apache 2.0 compatible
Model Details
| Property | Value |
|---|---|
| Base Model | Qwen3.5-4B |
| Fine-tuning | LoRA |
| Framework | Hugging Face Transformers |
| Training Library | Unsloth |
| Quantization | 4-bit (bnb) |
| Language | English |
| Developer | Anmol Dhandhania |
Benchmark Results
| Benchmark | Score |
|---|---|
| HellaSwag | 54.58 |
| HellaSwag (Norm) | 72.80 |
| PIQA | 78.13 |
| PIQA (Norm) | 78.45 |
| ARC-Easy | 80.13 |
| ARC-Easy (Norm) | 76.52 |
| Winogrande | 70.56 |
| TruthfulQA MC2 | 45.85 |
| MMLU | 73.34 |
MMLU Breakdown
| Category | Score |
|---|---|
| Humanities | 65.14 |
| Social Sciences | 82.94 |
| STEM | 71.99 |
| Other | 77.60 |
Improvements over v0.2
| Benchmark | Improvement |
|---|---|
| ARC-Easy | +2.27 |
| TruthfulQA MC2 | +3.33 |
| MMLU | +7.74 |
| Winogrande | +0.32 |
| PIQA | +0.16 |
These results indicate that targeted fine-tuning on TruthfulQA improved both factual accuracy and general reasoning performance across multiple evaluation tasks.
Intended Use
ANMOLGPT is designed for:
- Conversational AI
- Question Answering
- Reasoning Tasks
- Educational Applications
- Research Assistance
- Coding Assistance
- AI Experimentation
Limitations
- Primarily trained for English.
- May still produce incorrect or hallucinated responses.
- Not intended for high-risk domains such as medical, legal, or financial decision-making without expert oversight.
Training
Base Model:
- Qwen3.5-4B
Fine-tuning:
- TruthfulQA
- LoRA
- Unsloth
- Hugging Face Transformers
- 4-bit Quantization
Version History
| Version | Description |
|---|---|
| v0.1 | Initial release based on Qwen2.5-3B |
| v0.2 | Migrated to Qwen3.5-4B |
| v0.3 | Truthfulness-focused fine-tuning with improved benchmark performance |
Roadmap
Future work includes:
- Continued Pre-training
- Multi-dataset Instruction Tuning
- DPO / Preference Optimization
- Coding Improvements
- GSM8K
- HumanEval
- BBH
- MATH
- Long-context Support
Citation
@software{anmolgpt_v03,
title={ANMOLGPT-4B-v0.3},
author={Anmol Dhandhania},
year={2026},
version={0.3},
url={https://huggingface.co/anmoldhandhania93/ANMOLGPT-4B-v0.3}
}
Acknowledgements
ANMOLGPT is built upon the excellent work of:
- Alibaba Qwen Team
- Hugging Face
- Unsloth
- EleutherAI LM Evaluation Harness
- TruthfulQA authors
- Open-source AI community
If you find this model useful, consider ⭐ starring the repository and sharing your feedback. Contributions and suggestions are always welcome.
- Task
- Text Generation
- Type
- qwen3_5
- License
- apache-2.0
- Languages
- en
- Created
- Aug 4, 2026
- Updated
- Aug 6, 2026