---
license: llama2
base_model: meta-llama/Llama-2-13b-hf
pipeline_tag: text-generation
language:
  - en
tags:
  - llama-2
  - text-generation
  - rl-mpq
  - mixed-precision
  - quantization
  - fake-quantization
  - conservative
library_name: transformers
datasets:
  - wikitext
widget:
  - text: "The capital of France is"
---

# Llama 2 13B — RL-MPQ Conservative

Standalone **RL-MPQ** (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the
**Conservative** scenario — a quantized variant of
[meta-llama/Llama-2-13b-hf](https://huggingface.co/meta-llama/Llama-2-13b-hf).

| Field | Value |
|-------|-------|
| **Base model** | [meta-llama/Llama-2-13b-hf](https://huggingface.co/meta-llama/Llama-2-13b-hf) |
| **Scenario** | Conservative |
| **Avg bits / weight** | 5.2 |
| **Compression vs FP16** | 3.0769× |
| **WikiText-2 PPL** | 4.4663 |
| **Layers** | 40 |
| **Bit distribution** | `{'4': 28, '8': 12}` |
| **Format** | Fake-quant FP16 + `rlmpq_policy.json` |

**Collection:** [RL-MPQ — Llama 2 13B](https://huggingface.co/collections/AvoCahDoe/rl-mpq-llama-2-13b-6a2b04386dd0303714a19782) — all five scenarios for Llama 2 13B.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/llama-2-13b-rlmpq-conservative"

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

## Other Llama 2 13B scenarios

| Scenario | Avg bits | Compression | WikiText-2 PPL |
|----------|----------|-------------|----------------|
| [Aggressive](https://huggingface.co/AvoCahDoe/llama-2-13b-rlmpq-aggressive) | 3.75 | 4.2667x | 4.5724 |
| [Balanced](https://huggingface.co/AvoCahDoe/llama-2-13b-rlmpq-balanced) | 4.4 | 3.6364x | 4.4797 |
| [Extreme Survival](https://huggingface.co/AvoCahDoe/llama-2-13b-rlmpq-extreme-survival) | 2.775 | 5.7658x | 6.1148 |
| [High Fidelity](https://huggingface.co/AvoCahDoe/llama-2-13b-rlmpq-high-fidelity) | 6.7 | 2.3881x | 4.4313 |

Grouped archive (all scenarios in one repo):
[AvoCahDoe/llama-2-13b-rlmpq](https://huggingface.co/AvoCahDoe/llama-2-13b-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_llama_2_13b_conservative_2026,
  title  = {RL-MPQ Conservative: Llama 2 13B Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/llama-2-13b-rlmpq-conservative}
}
```