dhlak/llama-3.1-8b-alpaca-lora

Llama-3.1-8B LoRA - Alpaca Fine-tune

A LoRA adapter for Llama-3.1-8B fine-tuned on the Alpaca Cleaned dataset for instruction following.

Model Details

• Base Model: unsloth/Llama-3.1-8B
• Model Type: LoRA adapter (PEFT)
• Language: English
• License: Llama 3.1 Community License
• Fine-tuning Method: Supervised Fine-Tuning (SFT) with LoRA

LoRA Configuration

Parameter	Value
r (rank)	16
alpha	16
dropout	0
target_modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training Details

Parameter	Value
Dataset	yahma/alpaca-cleaned
Training Samples	51,760
Training Steps	809
Final Loss	1.20
Training Time	~1.7 hours
Precision	bf16
Peak GPU Memory	37.6 GB

Evaluation Results

Evaluated on IFEval (Instruction Following Evaluation):

Metric	Score
Prompt-level Strict Accuracy	22.4%
Prompt-level Loose Accuracy	22.4%
Instruction-level Strict Accuracy	36.7%
Instruction-level Loose Accuracy	36.7%

Usage

With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/Llama-3.1-8B",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("unsloth/Llama-3.1-8B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "YOUR_USERNAME/YOUR_MODEL_NAME")

# Generate
messages = [{"role": "user", "content": "Explain quantum computing in simple terms."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Unsloth (Faster)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="YOUR_USERNAME/YOUR_MODEL_NAME",
    max_seq_length=2048,
    load_in_4bit=True,  # or False for full precision
)
FastLanguageModel.for_inference(model)

# Generate
messages = [{"role": "user", "content": "Explain quantum computing in simple terms."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

• This is a base model fine-tuned on Alpaca, not an instruct model. Response quality may vary.
• The model inherits limitations from the base Llama-3.1-8B model.
• Not suitable for production use without further evaluation and safety testing.

Framework Versions

• PEFT: 0.18.1
• Transformers: 4.x
• Unsloth: latest