🎯 Zen Alta 1-3B (Phase 2 DPO LoRA Adapter)
This repository contains the LoRA adapter weights (132.5 MB) from Phase 2 Direct Preference Optimization (DPO) of the Zen Alta project.
The adapter was trained for 800 steps on top of the 24-layer pruned base model ZenithLLM/ZenAlta-1-3B-Pruned to engrave a casual, witty conversational persona with targeted corporate-AI unlearning.
🚀 Loading and Merging with Transformers / PEFT
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
BASE_MODEL = "ZenithLLM/ZenAlta-1-3B-Pruned"
LORA_ADAPTER = "ZenithLLM/ZenAlta-1-3B-Phase2"
tokenizer = AutoTokenizer.from_pretrained(LORA_ADAPTER)
base_model = AutoModelForCausalLM.from_pretrained(
BASE_MODEL,
torch_dtype=torch.float16,
device_map="auto"
)
# Attach LoRA adapter
model = PeftModel.from_pretrained(base_model, LORA_ADAPTER)
# (Optional) Merge into standalone weights:
merged_model = model.merge_and_unload()
🔗 Related Artifacts
- Ready-to-Run GGUF (Q4_K_M): ZenithLLM/ZenAlta-1-3B-Phase2-GGUF
- Speculative Draft Model: ZenithLLM/ZenAlta-Draft
- Pruned Base Model: ZenithLLM/ZenAlta-1-3B-Pruned