LevelUp AI: QLoRA adapter for Llama-3.2-3B-Instruct
1 min readupdated
A new QLoRA adapter (QLoRA: an efficient low-precision fine-tuning method for large language models) equips the base Llama-3.2-3B-Instruct model with seven distinct coaching personas for a gamified self-development app. Trained on 76,616 curated examples, the release contains only the adapter weights, which need to be loaded on top of the base model. This solution suits developers building interactive assistant bots with unique conversational styles.
Key features
- Seven unique coaching personas for diverse interaction scenarios.
- Trained via QLoRA in 4-bit NF4 format using the transformers and peft libraries.
- Optimized target modules for q, k, v, o, gate, up, and down projections.
- Documented evaluation on 200 held-out examples showing improved ROUGE metrics.
What you can do
- Integrate the adapter into a Telegram bot or web app for interactive coaching.
- Leverage different personas for gamified self-improvement and user motivation tasks.
Example
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "unsloth/Llama-3.2-3B-Instruct"
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "veeravenkatsai/levelup-qlora-llama3.2-3b")
tok = AutoTokenizer.from_pretrained(base)Read more
DetailsBenchmarks (on 200 examples): ROUGE-L increased from 0.1570 to 0.2665 (+70% relative improvement).
Training parameters: 3 epochs, learning rate 2e-4, batch size 16, maximum sequence length 512.
Limitations: the adapter has not undergone specialized safety tuning for medical, financial, or psychological advice.










Comments