---
language:
  - en
license: gemma
library_name: transformers
pipeline_tag: text-generation
base_model: google/gemma-4-31B-it
widget:
  - text: "gemma"
    output:
      url: gemma.webp
tags:
  - text-generation
  - security
  - red-team
  - telemetry
  - local-first
  - safetensors
  - litert-lm
  - dpm
---

<div align="center">
  <img src="gemma.webp" width="180px" style="border-radius: 10px; margin-bottom: 20px;">
  <h1>google/gemma-4-31B-it <br><sub>LiteRT-LM Optimized</sub></h1>
  <p><i>Deterministic Projection Memory (DPM) Artifact for Security Telemetry</i></p>
</div>

---

### 🟢 Overview
This repository contains a specialized **LiteRT-LM conversion** of `google/gemma-4-31B-it`. It is engineered for local-first **DPM//BENCH** experiments, specifically targeting long-horizon incident narratives and red-team traces.

> **Objective:** Package the base instruction model into a runtime format for deterministic projection memory experiments, ensuring that append-only event logs map to a consistent structured memory surface.

---

### 🛠️ Conversion Architecture
The conversion utilizes the latest LiteRT-LM stack, requiring specific flags to handle the Gemma 4 per-layer embedding structure.

<details>
<summary><b>View Conversion Script</b></summary>

```bash
python -m litert_torch.generative.export_hf \
    --model /path/to/google/gemma-4-31B-it \
    --output_dir /path/to/out/gemma-4-31B-it-litert-lm \
    --externalize_embedder True \
    --single_token_embedder True \
    --experimental_lightweight_conversion True \
    --bundle_litert_lm True \
    --task text_generation
```
</details>

**Critical Flags for Compatibility:**
* **--externalize_embedder True**: Essential for per-layer embedding paths.
* **--experimental_lightweight_conversion True**: Prevents runtime artifact corruption.
* **--bundle_litert_lm True**: Packages tokenizer and templates into the `.litertlm` artifact.

---

### 💻 Infrastructure Requirements

| Requirement | Specification | Context |
| :--- | :--- | :--- |
| **RAM** | 128 GB+ | Minimum for 31B conversion overhead |
| **Disk Space** | 500 GB | Workspace for intermediate FlatBuffer assets |
| **Storage Type** | NVMe SSD | Crucial for large model serialization |
| **Inference** | Apple Silicon / GPU | 31B is unsuitable for fast CPU-only DPM |

---

### 🔍 Validation Protocol
For a successful **DPM//BENCH** run, the artifact must maintain byte-stability. Ensure the following conditions are met:

1. **Integrity:** LiteRT-LM binary successfully parses the `.litertlm` bundle.
2. **Determinism:** At `temp 0` and a fixed seed, repeated projection calls must yield identical memory-surface bytes.
3. **Format:** JSON-only prompts must satisfy schema constraints under high-compression DPM tests.

---

### ⚠️ Implementation Boundaries
* **Intended Use:** Security incident summarization, telemetry trace compression, and blue-team event reasoning.
* **Non-Intended Use:** This is not a standalone decision-making system. It is a projection tool. All outputs require human review and replay-validation in high-stakes environments.

---

<div align="right">
  <p><i>Base Model: google/gemma-4-31B-it</i></p>
</div>