Llama 3 8B — RL-MPQ Balanced
Standalone RL-MPQ (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the Balanced scenario — a quantized variant of meta-llama/Meta-Llama-3-8B.
| Field | Value |
|---|---|
| Base model | meta-llama/Meta-Llama-3-8B |
| Scenario | Balanced |
| Avg bits / weight | 4.5 |
| Compression vs FP16 | 3.5556× |
| WikiText-2 PPL | 6.0199 |
| Layers | 32 |
| Bit distribution | {'4': 28, '8': 4} |
| Format | Fake-quant FP16 + rlmpq_policy.json |
Collection: RL-MPQ — Llama 3 8B — all five scenarios for Llama 3 8B.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "AvoCahDoe/llama-3-8b-rlmpq-balanced"
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo)
Other Llama 3 8B scenarios
| Scenario | Avg bits | Compression | WikiText-2 PPL |
|---|---|---|---|
| High Fidelity | 6.875 | 2.3273x | 5.8133 |
| Conservative | 5.25 | 3.0476x | 5.9652 |
| Aggressive | 3.7188 | 4.3025x | 6.6793 |
| Extreme Survival | 2.875 | 5.5652x | 32.393 |
Grouped archive (all scenarios in one repo): AvoCahDoe/llama-3-8b-rlmpq
Method
- Phase 3 — PPO agent assigns per-layer bit widths under the Balanced reward target.
- Phase 4 — Policy replayed on real weights; WikiText-2 perplexity validates quality.
- 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
@misc{rlmpq_llama_3_8b_balanced_2026,
title = {RL-MPQ Balanced: Llama 3 8B Mixed-Precision Quantization},
author = {AvoCahDoe},
year = {2026},
url = {https://huggingface.co/AvoCahDoe/llama-3-8b-rlmpq-balanced}
}