🦅 Falcon-RW-1B Fine-Tuned on Dolly 15K using QLoRA

This model is a fine-tuned version of tiiuae/falcon-rw-1b on the Dolly 15K dataset using QLoRA. The training was done using the LLaMA-Factory framework on a Google Colab T4 GPU.

The model is designed for instruction-following, fine-tuned with 4-bit quantization and LoRA adapters, making it memory-efficient and suitable for lightweight deployment.

🛠️ Model Details

Model type: Causal language model
Base model: Falcon-RW-1B
Finetuned from: tiiuae/falcon-rw-1b
Dataset: Dolly 15K
Language: English
Quantization: 4-bit (QLoRA)
Adapters: LoRA
Trained using: LLaMA-Factory
License: Apache 2.0
Hardware: Google Colab (Tesla T4, 15 GB VRAM)
Training time: ~2.3 hours
Epochs: 3

📦 How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("SindhuraSriram/falcon-rw-1b-dolly-qlora")
tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-rw-1b")

prompt = "Explain quantum computing to a 10-year-old."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)

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

📊 Training Summary

Metric	Value
GPU Used	Tesla T4 (Colab)
VRAM	~11.5 GB
Time Taken	~2.3 hours
Batch Size	2
Epochs	3

✅ Intended Use

Direct Use

Instruction-following tasks
Text generation
Educational Q&A
Lightweight deployment via LoRA adapters

Out-of-Scope Use

Tasks requiring up-to-date real-world knowledge
Sensitive applications with fairness or safety concerns

🧠 Bias, Risks, and Limitations

This model inherits limitations and potential biases from:

The base Falcon-RW-1B model
The Dolly 15K dataset (derived from filtered ChatGPT outputs)

It may reflect social, cultural, or demographic biases present in its training data. It is not guaranteed to be safe for all use cases.

🧾 Citation

If you use this model, please cite the original Falcon and Dolly authors, and optionally credit this model as:

APA:
Sriram, S. (2025). Falcon-RW-1B Fine-Tuned on Dolly 15K with QLoRA. HuggingFace. https://huggingface.co/SindhuraSriram/falcon-rw-1b-dolly-qlora

🙋‍♀️ Contact

For any questions or support:

Sindhura Sriram
📧 [email protected]
🔗 LinkedIn
🌐 Portfolio

🌍 Environmental Impact

Factor	Estimate
GPU Type	Tesla T4
Time Used	~2.3 hours
Platform	Google Colab
Region	Unknown
CO₂ Estimate	Low (educational use)

🔗 Model Sources

💡 Acknowledgements

Thanks to HuggingFace, TII Falcon, Databricks, and LLaMA-Factory for enabling this open-source experiment.