---
license: gemma
base_model: google/gemma-2-9b
pipeline_tag: text-generation
language:
  - en
tags:
  - gemma
  - text-generation
  - rl-mpq
  - mixed-precision
  - quantization
  - fake-quantization
  - conservative
library_name: transformers
datasets:
  - wikitext
widget:
  - text: "The capital of France is"
---

# Gemma 2 9B — RL-MPQ Conservative

Standalone **RL-MPQ** (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the
**Conservative** scenario — a quantized variant of
[google/gemma-2-9b](https://huggingface.co/google/gemma-2-9b).

| Field | Value |
|-------|-------|
| **Base model** | [google/gemma-2-9b](https://huggingface.co/google/gemma-2-9b) |
| **Scenario** | Conservative |
| **Avg bits / weight** | 5.1429 |
| **Compression vs FP16** | 3.1111× |
| **WikiText-2 PPL** | 116.5244 |
| **Layers** | 42 |
| **Bit distribution** | `{'4': 30, '8': 12}` |
| **Format** | Fake-quant FP16 + `rlmpq_policy.json` |

**Collection:** [RL-MPQ — Gemma 2 9B](https://huggingface.co/collections/AvoCahDoe/rl-mpq-gemma-2-9b-6a2b1d9443205eab61ff447b) — all five scenarios for Gemma 2 9B.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/gemma-2-9b-rlmpq-conservative"

model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo)
```

## Other Gemma 2 9B scenarios

| Scenario | Avg bits | Compression | WikiText-2 PPL |
|----------|----------|-------------|----------------|
| [Aggressive](https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq-aggressive) | 3.6667 | 4.3636x | 162.8437 |
| [Balanced](https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq-balanced) | 4.2857 | 3.7333x | 127.0798 |
| [Extreme Survival](https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq-extreme-survival) | 2.7857 | 5.7436x | 424.7991 |
| [High Fidelity](https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq-high-fidelity) | 7.0476 | 2.2703x | 104.8098 |

Grouped archive (all scenarios in one repo):
[AvoCahDoe/gemma-2-9b-rlmpq](https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq)

## Method

1. **Phase 3** — PPO agent assigns per-layer bit widths under the Conservative reward target.
2. **Phase 4** — Policy replayed on real weights; WikiText-2 perplexity validates quality.
3. **Export** — Fake-quantized FP16 weights compatible with Hugging Face Transformers.

## Files

| File | Description |
|------|-------------|
| `config.json` | Llama architecture + RL-MPQ metadata |
| `model.safetensors` | Fake-quantized weights |
| `rlmpq_policy.json` | Per-layer bit-width policy |
| `rlmpq_metrics.json` | Validation & PPL summary |

## Citation

```bibtex
@misc{rlmpq_gemma_2_9b_conservative_2026,
  title  = {RL-MPQ Conservative: Gemma 2 9B Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/gemma-2-9b-rlmpq-conservative}
}
```