Train Zen Max identity using efficient QLoRA fine-tuning.

• Cloud Training: All training happens on HuggingFace - no local downloads
• INT4 Base Model: K2 already quantized to INT4 (~370GB)
• QLoRA Efficiency: LoRA adapters on INT4 model (multi-GPU training)
• LoRA Adapters: Only trains adapters (~100MB) not full model
• Auto-Upload: Adapters pushed directly to zenlm/zen-max

• Model: zenlm/zen-max-base
• Size: 671B parameters (384 experts, 8 active)
• BF16 Weights: ~1.3TB (full precision)
• INT4 Weights: ~370GB (quantized on HuggingFace)
• Training Method: QLoRA on INT4 model (requires multi-GPU)

• GPU: 4x A100 80GB or 8x A100 40GB (provided by Space)
• Model Size: ~370GB (INT4 quantized, 62 shards)
• VRAM Usage: ~500GB total (370GB model + ~130GB activations)
• Training Time: 4-8 hours for 1000 steps

• Rank: 16 (adjustable 4-64)
• Alpha: 32
• Dropout: 0.05
• Target Modules: All attention and MLP layers

• Source: zenlm/zen-identity-dataset
• Content: Zen persona, values, and conversational patterns
• Size: Curated high-quality identity examples

LoRA Adapters: Uploaded to zenlm/zen-max

• Adapter weights: ~100MB
• Compatible with zen-max-base model
• Preserves all reasoning capabilities
• Adds Zen identity and values

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model (can use 4-bit for inference too)
base_model = AutoModelForCausalLM.from_pretrained(
    "zenlm/zen-max-base",
    device_map="auto",
    load_in_4bit=True
)

# Load Zen adapters
model = PeftModel.from_pretrained(base_model, "zenlm/zen-max")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-max")

# Inference with Zen identity
messages = [{"role": "user", "content": "Tell me about yourself"}]
response = model.chat(tokenizer, messages, thinking_budget=128000)

1. No Downloads: Train on cloud, never download 1TB model locally
2. Efficient: QLoRA uses 4-bit quantization for minimal memory
3. Modular: Adapters can be loaded on top of any zen-max-base checkpoint
4. Practical: Inference can also use 4-bit for consumer hardware

• Base Model: https://huggingface.co/zenlm/zen-max-base
• Output Repo: https://huggingface.co/zenlm/zen-max
• Organization: https://huggingface.co/zenlm
• Website: https://zenlm.org

Zen AI: Clarity Through Intelligence