anmoldhandhania93

ANMOLGPT-4B-v0.2

README

License: apache-2.0

Overview

ANMOLGPT-4B-v0.2 is an instruction-following language model built by fine-tuning Qwen3.5-4B using QLoRA with Unsloth Studio.

This release builds upon the foundation established by ANMOLGPT-3B-v0.1, introducing a more capable base model while preserving the project's focus on reproducible training, transparent evaluation, and open-source development.

The objective of v0.2 is to improve reasoning, instruction following, and overall language understanding while providing a stronger foundation for future releases.

Base Model

Model: Qwen3.5-4B
Fine-tuning Method: QLoRA
Framework: Unsloth Studio
Export Format: Hugging Face Transformers

Training Details

Dataset

Databricks Dolly 25K

Training Configuration

Table with columns: Parameter, Value
Parameter	Value
Base Model	Qwen3.5-4B
Fine-tuning	QLoRA
Quantization	4-bit
Framework	Unsloth Studio
Optimizer	AdamW
Precision	4-bit

Benchmark Results

Benchmark Comparison

Table with columns: Benchmark, ANMOLGPT-3B-v0.1, ANMOLGPT-4B-v0.2
Benchmark	ANMOLGPT-3B-v0.1	ANMOLGPT-4B-v0.2
HellaSwag (acc_norm)	73.21	73.25
PIQA	77.97	77.97
ARC-Easy	77.95	77.86
Winogrande	70.24	70.24
TruthfulQA MC2	42.52	42.52
MMLU	65.59

Example Usage

python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "anmoldhandhania93/ANMOLGPT-4B-v0.2"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto"
)

prompt = "Explain how reinforcement learning differs from supervised learning."

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Intended Uses

ANMOLGPT-4B-v0.2 is suitable for:

Conversational AI
Learning and experimentation
Prompt engineering
Educational applications
Software development assistance
General-purpose text generation

Limitations

This model is an experimental open-source release.

Limitations include:

May produce inaccurate or fabricated information.
Performance has not yet been fully benchmarked across all standard evaluations.
Not intended for safety-critical or production use without additional validation.
Outputs should be reviewed before use in professional environments.

Roadmap

v0.2

✅ Upgraded to Qwen3.5-4B
✅ Improved training pipeline
🔄 Comprehensive benchmark evaluation

v0.5

Mixed high-quality instruction datasets
Longer training schedule
Improved reasoning and coding capabilities

v1.0

Domain-specific continued pre-training
Advanced instruction tuning
Human evaluation
Retrieval-Augmented Generation (RAG)
Tool calling
Production-ready deployment

Project Goals

The long-term goal of ANMOLGPT is to explore efficient and transparent development of open-source language models while documenting the complete engineering lifecycle:

Training
Fine-tuning
Benchmarking
Deployment
Continuous improvement

Acknowledgements

ANMOLGPT is built using the outstanding work of the open-source AI community.

Special thanks to:

Qwen Team
Unsloth AI
Hugging Face
EleutherAI
Databricks (Dolly Dataset)

Citation

bibtex
@misc{anmolgpt2026,
  title={ANMOLGPT-4B-v0.2},
  author={Anmol Dhandhania},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/anmoldhandhania93/ANMOLGPT-4B-v0.2}
}

Future Work

Upcoming releases will focus on:

Higher-quality instruction datasets
Continued pre-training for domain adaptation
Stronger reasoning capabilities
Coding improvements
Comprehensive benchmark comparisons
Efficient inference optimization

Feedback, issues, and contributions are always welcome.

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