⚡ Veloxy-Flash-15B-LoRA: Faz 8 Tool-Call Master Adapter ⚡

Official Faz 8 Tool-Call Master LoRA Weights | Rank 16, Alpha 32 | Converged Loss: 0.2799 | 38 Native Tools & Java 21 LTS

---

📌 Overview

This repository contains the official Faz 8: Tool-Call Master Low-Rank Adaptation (LoRA) weights trained for Veloxy-Flash-15B.

Trained across 4,342 curated enterprise-grade multi-turn samples on an NVIDIA RTX A6000 48GB GPU cluster, the adapter achieved a record convergence training loss of 0.2799 at step 160.

🌟 Adapter Specializations:

• 38 Native Full-OS & Browser Tools: Directly emits <tool_call> blocks with strict JSON arguments.
• Intelligent Unknown Tool Discovery: Dynamically detects unfamiliar tool names and runs schema introspection / --help routines.
• Zero-Refusal Cybersecurity Analysis: Deep technical exploit analysis (ASLR/DEP evasion, GOT overwrite, ROP chains) and SAST code auditing.
• Java 21 LTS & JVM Internals: Project Loom Virtual Threads, ZGC/Shenandoah/G1 GC mechanics, and JIT C1/C2 tuning.
• 2026 Temporal Anchor: Contemporary 2026 technical stack alignment without legacy temporal drift.

---

⚙️ LoRA Hyperparameters

Parameter	Value
Base Model	zveloxy/Veloxy-Flash-15B
LoRA Rank ($r$)	16
LoRA Alpha ($\alpha$)	32
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (All 7 linear projections)
Training Precision	bfloat16 (Paged AdamW 8-bit optimizer)
Convergence Loss	0.2799 (Step 160)

---

🚀 Usage & Base Model Merging

A. Dynamic Inference via PEFT

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "zveloxy/Veloxy-Flash-15B"
lora_model_id = "zveloxy/Veloxy-Flash-15B-LoRA"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Attach Faz 8 LoRA Adapter
model = PeftModel.from_pretrained(base_model, lora_model_id)

messages = [
    {"role": "system", "content": "You are Veloxy-Flash-15B, an autonomous reasoning agent."},
    {"role": "user", "content": "List the top 5 memory-consuming processes using bash_exec."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

B. Permanent Merge into Standalone Safetensors

To fuse the LoRA weights permanently into the base model and export standalone BF16 weights:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "zveloxy/Veloxy-Flash-15B"
lora_model_id = "zveloxy/Veloxy-Flash-15B-LoRA"
output_dir = "./veloxy-flash-15b-merged"

print("[*] Loading base model...")
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="cpu",
    trust_remote_code=True
)

print("[*] Merging Faz 8 LoRA adapter...")
lora_model = PeftModel.from_pretrained(base_model, lora_model_id)
merged_model = lora_model.merge_and_unload()

print(f"[*] Saving merged model to {output_dir}...")
merged_model.save_pretrained(output_dir, max_shard_size="5GB", safe_serialization=True)

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
tokenizer.save_pretrained(output_dir)
print("[+] Successfully merged and saved!")

---

🌐 Ecosystem

• Base Model: zveloxy/Veloxy-Flash-15B
• Flagship 32B: zveloxy/Veloxy-Titan-32B
• Multimodal 32B: zveloxy/Veloxy-Titan-32B-Vision

---

📄 License

Licensed under the Apache License 2.0.