Supervised fine-tune of TinyLlama/TinyLlama-1.1B-Chat-v1.0 on the Stanford Alpaca instruction dataset using LoRA (4-bit QLoRA training). The LoRA adapter has been merged into the base weights, so this repo is a full, ready-to-run model — no PEFT needed.

NLP Assignment 4 — Track 1 (LLM Fine-Tuning), Option B (Multi-Dataset SFT). This is the best-performing configuration (Alpaca, trial 4) selected from 10 trials across two datasets (Alpaca and Dolly-15k).

Results (10 manual test prompts, vs. ChatGPT-4o gold answers)

Model	BLEU	BERTScore F1
Baseline TinyLlama	3.7116	0.8732
This model (Alpaca SFT)	4.7111	0.8710

The Alpaca fine-tune improved BLEU by ~27% over the baseline.

Training configuration

Hyperparameter	Value
Method	QLoRA (4-bit NF4)
LoRA rank (r)	32
LoRA alpha	64
Target modules	q_proj, v_proj
LoRA dropout	0.05
Learning rate	3e-4
Epochs	2
Dataset	tatsu-lab/alpaca (2,000 samples)
Hardware	Kaggle Tesla T4

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "AnasTabba/tinyllama-alpaca-sft"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

prompt = "<|user|>\nExplain what machine learning is in simple terms.</s>\n<|assistant|>\n"
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200, temperature=0.7, do_sample=True)
print(tok.decode(out[0], skip_special_tokens=True))