Llama 3.1 8B — RL-MPQ Conservative
Standalone RL-MPQ (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the Conservative scenario — a quantized variant of meta-llama/Llama-3.1-8B.
| Field | Value |
|---|---|
| Base model | meta-llama/Llama-3.1-8B |
| Scenario | Conservative |
| Avg bits / weight | 5.25 |
| Compression vs FP16 | 3.0476× |
| WikiText-2 PPL | 5.7182 |
| Layers | 32 |
| Bit distribution | {'4': 22, '8': 10} |
| Format | Fake-quant FP16 + rlmpq_policy.json |
Collection: RL-MPQ — Llama 3.1 8B — all five scenarios for Llama 3.1 8B.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "AvoCahDoe/llama-3-1-8b-rlmpq-conservative"
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo)
Other Llama 3.1 8B scenarios
| Scenario | Avg bits | Compression | WikiText-2 PPL |
|---|---|---|---|
| Aggressive | 3.5938 | 4.4522x | 6.4787 |
| Balanced | 4.375 | 3.6571x | 5.7761 |
| Extreme Survival | 2.6875 | 5.9535x | 34.8686 |
| High Fidelity | 6.75 | 2.3704x | 5.5974 |
Grouped archive (all scenarios in one repo): AvoCahDoe/llama-3-1-8b-rlmpq
Method
- Phase 3 — PPO agent assigns per-layer bit widths under the Conservative 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_1_8b_conservative_2026,
title = {RL-MPQ Conservative: Llama 3.1 8B Mixed-Precision Quantization},
author = {AvoCahDoe},
year = {2026},
url = {https://huggingface.co/AvoCahDoe/llama-3-1-8b-rlmpq-conservative}
}