---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B
pipeline_tag: text-generation
language:
  - en
tags:
  - qwen2
  - text-generation
  - rl-mpq
  - mixed-precision
  - quantization
  - fake-quantization
  - aggressive
library_name: transformers
datasets:
  - wikitext
widget:
  - text: "The capital of France is"
---

# Qwen 2.5 7B — RL-MPQ Aggressive

Standalone **RL-MPQ** (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the
**Aggressive** scenario — a quantized variant of
[Qwen/Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B).

| Field | Value |
|-------|-------|
| **Base model** | [Qwen/Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B) |
| **Scenario** | Aggressive |
| **Avg bits / weight** | 3.1429 |
| **Compression vs FP16** | 5.0909× |
| **WikiText-2 PPL** | 9.3678 |
| **Layers** | 28 |
| **Bit distribution** | `{'3': 24, '4': 4}` |
| **Format** | Fake-quant FP16 + `rlmpq_policy.json` |


## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/qwen2-5-7b-rlmpq-aggressive"

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

## Other Qwen 2.5 7B scenarios

| Scenario | Avg bits | Compression | WikiText-2 PPL |
|----------|----------|-------------|----------------|
| [Balanced](https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq-balanced) | 3.3929 | 4.7158x | 8.9305 |
| [Conservative](https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq-conservative) | 3.6786 | 4.3495x | 8.4114 |
| [Extreme Survival](https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq-extreme-survival) | 2.4643 | 6.4928x | 497.4791 |
| [High Fidelity](https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq-high-fidelity) | 3.75 | 4.2667x | 8.208 |

Grouped archive (all scenarios in one repo):
[AvoCahDoe/qwen2-5-7b-rlmpq](https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq)

## Method

1. **Phase 3** — PPO agent assigns per-layer bit widths under the Aggressive 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_qwen2_5_7b_aggressive_2026,
  title  = {RL-MPQ Aggressive: Qwen 2.5 7B Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/qwen2-5-7b-rlmpq-aggressive}
}
```