ft-lab-tinyllama-1.1b-finetuning

FT-Lab: TinyLlama-1.1B Fine-Tuning Baselines (Full-FT / LoRA / QLoRA)

A minimal, fully reproducible fine-tuning stack for TinyLlama-1.1B,
providing clean baselines for Full Fine-Tuning, LoRA, and QLoRA.
Designed for Colab / T4 / A10 environments.

---

Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Sai1202HF/ft-lab-tinyllama-1.1b-finetuning"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

inputs = tokenizer("Explain LoRA.", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128)

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

Model Summary

This model is part of FT-Lab, a reproducible research toolkit for TinyLlama fine-tuning.
It serves as a baseline reference checkpoint for comparing Full-FT, LoRA, and QLoRA.

Although FT-Lab includes RAG and retrieval evaluation pipelines,
this uploaded checkpoint is not a RAG-optimized model.
It is intended for instruction tuning and fine-tuning studies.

This model is a fine-tuned variant of TinyLlama/TinyLlama-1.1B-Chat-v1.0,
optimized for small-GPU environments using:

• Full Fine-Tuning
• LoRA
• QLoRA

The goal is to provide a minimal, reproducible baseline for:

• instruction tuning
• RAG evaluation experiments
• small-scale research
• controlled ablation studies
• comparison of FT / LoRA / QLoRA behaviors

All training and evaluation scripts belong to FT-LLab,
designed for lightweight environments such as T4 or A10 GPUs.

Intended Use

This model is intended for:

• Educational understanding of fine-tuning pipelines
• Baseline research experiments
• Small-scale RAG + instruction tuning studies
• Method comparison of Full FT / LoRA / QLoRA

Not intended for high-risk domains (finance, healthcare, legal) without further evaluation and safeguards.

Training Procedure

• Base Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• Training Methods: Full Fine-Tuning / LoRA / QLoRA
• Hardware: T4 / A10 / Colab Pro
• Framework: PyTorch + HuggingFace Transformers
• Optimizer: AdamW

Hyperparameters and scripts are fully reproducible inside FT-Lab.

Dataset

All datasets in the public demo (toy_qa.jsonl, sample_eval.jsonl) are synthetic dummy datasets used solely to demonstrate the FT-Lab pipeline.

They do not represent meaningful real-world semantic content.

For real experiments, replace the dataset under data/ with your own or a public instruction-tuning dataset.

Evaluation

FT-Lab includes evaluation using:

• Exact Match (EM)
• Token-level accuracy
• BERTScore (optional)
• Custom RAG verification pipeline

These metrics support relative comparison between methods, not benchmark-grade scores.

Limitations

• Model size is small (1.1B), limiting reasoning and factual accuracy.
• Training data in the demo is synthetic.
• May hallucinate or produce incorrect content.
• Safety alignment is minimal.

Ethical Considerations

Use in safety-critical or high-risk settings requires:

• Additional evaluation
• Guardrails / filtering
• Human oversight

Citation

If you use FT-Lab or this model, please cite the repository:

https://github.com/REICHIYAN/ft_lab