TezzLLM Training Guide

This guide describes how to operate local training pipelines using pre-compiled binaries distributed under the TezzCorp Private Source License (TPSL v1.0).

Prerequisites

• OS: Windows 10/11 x64
• RAM: 256 MB minimum (512 MB recommended for multi-seeded training)
• CPU: Any standard x64 processor (more cores = faster parallel training)
• Disk: 200 MB free space
• Deployment License: TPSL v1.0

Step 1: Download & Extract Binaries

1. Navigate to the Download Portal.
2. Download the Pre-compiled Executables Package (tezzllm-tiny-v2-windows-x64.zip).
3. Extract the contents into a secure directory on your local machine.

The package contains:

• tezzllm_data.exe — Tokenizer utility
• tezzllm_v2_train.exe — Double-precision CPU trainer (vectorized)
• tezzllm_v2_merge.exe — Weight merger utility
• tezzllm_v2_engine.exe — Local chat engine
• build_v2.ps1 — Parallel training orchestrator script

Step 2: Prepare Training Data

Prepare raw plain UTF-8 text datasets containing your private domain knowledge.

• Place text files in the tezzllm_weights/ directory.
• Ensure files use the .txt extension (e.g. tezzllm_weights/corpus.txt).
• For best results, use clean, well-formed text.

Step 3: Tokenize the Dataset

Run the pre-compiled tokenizer binary in your PowerShell terminal:

.\tezzllm_data.exe

Expected output:

+---------------------------------------------------+
|  TezzLLM Data Processor v1.0 | Pure TezzNative   |
+---------------------------------------------------+

[] Reading tezzllm_weights/corpus.txt...

[OK] Read 187,672 bytes
[] Tokenizing (byte-level)...
[OK] Wrote 187672 tokens -> tezzllm_weights/tezzllm_train.bin

Step 4: Run Offline Training

Multi-Core Parallel Training (Recommended)

Run the Windows parallel training script:

.\build_v2.ps1

This training orchestrator script:

1. Detects available CPU cores.
2. Spawns parallel trainer instances (tezzllm_v2_train.exe) running under different random seeds.
3. Automatically merges trained weights using the federated merger utility tezzllm_v2_merge.exe upon completion.
4. Outputs the final combined weights file to tezzllm_weights/tezzllm_v2_merged.tezw.

Single Seed Training

To run a single training sequence manually:

.\tezzllm_v2_train.exe --seed 42 --threads 4

Step 5: Adjust Settings in Command Arguments

Under the TPSL license, raw compiler source code is protected. Rather than recompiling the source code, training parameters are adjusted directly in execution commands or passed via parameter flags:

• --epochs <value>: Set total training cycles.
• --lr <value>: Modify learning rate.
• --seed <value>: Supply a specific random seed.
• --threads <value>: Allocate specific CPU thread count.

If custom optimizations, architectures, or features are required, TezzCorp delivers official updates based on enterprise demand or error logs.

Step 6: Monitor Convergence

Watch the training progress logs to monitor model convergence. The trainer outputs step losses:

[*] Training: 20 epochs...
[step 500/14966] loss=4.823    ← Loss should decrease over time
[step 1000/14966] loss=4.312
[step 1500/14966] loss=3.891
[step 2000/14966] loss=3.654
...
[ep 1/20] loss=4.234
[ep 2/20] loss=3.876

• Loss Guide:

- 4.0 = Extracting core pattern syntax

- 2.5 = Understanding local context

- 1.2 = High convergence (model has memorized data patterns)

Step 7: Local Inference

Once weights are averaged and compiled, start the chat interface:

.\tezzllm_v2_engine.exe

Step 8: Federated Weight Submission

To help improve the shared base model, submit the resulting 8.66 MB file:

tezzllm_weights/tezzllm_v2_merged.tezw

to the Weights Contribution Portal.

Only mathematical weight matrices are uploaded; raw datasets remain secure inside your local server limits.