Velthur/VAI-0.2-Distill-32B

VAI-0.2 (Velthur Adaptive Intelligence)

By Velthur Intelligence Lab (VIL) | Pakistan

VAI-0.2 is a locally adapted, 32-billion parameter deliberative reasoning model. Built on top of deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, VAI-0.2 is engineered to solve the catastrophic "Trace Collapse" vulnerability inherent in standard instruction fine-tuning.

The Engineering

Standard instruction fine-tuning on trace-free target data acts as a negative regularizer on deliberative models, forcing them to suppress their <think> manifolds to minimize cross-entropy loss.

To deploy frontier reasoning locally under extreme hardware constraints, Velthur Intelligence Lab utilized a Trace-Augmented Self-Distillation pipeline. We forced the frozen base model to synthesize explicit intermediate logic for our target formatting data prior to QLoRA adaptation.

Result: VAI-0.2 adheres to strict enterprise output formats (like deterministic JSON) while perfectly preserving its 32-billion parameter cognitive architecture.

Infrastructure

• Compute: Trained on a single NVIDIA A100 (40GB) on Google Colab.
• Precision: 4-bit NormalFloat (NF4) base quantization with BF16 Adapter training.
• Methodology: Trace-Augmented QLoRA + Test-Time Autonomous Verification.

Usage

VAI-0.2 is distributed as a PEFT/LoRA adapter. It must be loaded on top of the base DeepSeek-R1-Distill-Qwen-32B model.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
adapter_id = "Velthur/VAI-0.2-Distill-32B"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, load_in_4bit=True)

# Load the VAI-0.2 Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)