xxue752/llama3.2-caiti-full

Llama 3.2 3B Instruct — CaiTI (full merged weights)

This repository contains standalone merged weights for a Llama 3.2 3B Instruct model fine-tuned for the CaiTI (Conversational AI Therapist Interface) setup.

What this is

• Base: meta-llama/Llama-3.2-3B-Instruct
• Adaptation: CaiTI LoRA adapters for Task 1, Task 2, and Task 3, linearly combined into a single adapter, then merged into the base with merge_and_unload() so you do not need PEFT at inference.
• LoRA-only checkpoint (smaller download if you already have the base): xxue752/llama3.2-caiti

Loading (Transformers)

Use a recent transformers build and a GPU with enough memory for 3B in bf16/fp16.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "xxue752/llama3.2-caiti-full"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

If the model is gated or you use a private token:

tokenizer = AutoTokenizer.from_pretrained(model_id, token=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    token=True,
)

How the merge was produced

Merged locally with CaiTI/llama-3.2-3b-finetune/export_merged_full_model.py: load base + adapter from caiti-llama32-3b-ultimate, then merge_and_unload(), save as model.safetensors (bf16 on GPU).

Training details and prompts

Instruction formats and training code live in the source project under llama-3.2-3b-finetune/ (e.g. llama_finetune_caiti.py, llama32_3b_finetune_task3.py).

License

This model is derived from Llama 3.2; use is subject to the Meta Llama 3.2 Community License. Do not use this model if you have not accepted the base model’s license on Hugging Face.