RL-MPQ repo index (README + config.json) — 2026-06-11T15:57:56.128906

README.md

@@ -2,20 +2,36 @@
 license: llama2
 base_model: meta-llama/Llama-2-7b-hf
 pipeline_tag: text-generation
+language:
+  - en
 tags:
+  - llama
+  - text-generation
   - rl-mpq
   - mixed-precision
   - quantization
-  - text-generation
+  - fake-quantization
+  - llama-2
 library_name: transformers
+datasets:
+  - wikitext
 ---
 
-# RL-MPQ — LLAMA-2-7B (all scenarios)
+# Llama 2 7B — RL-MPQ Quantized (Thesis Release)
+
+**Quantized variant of [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf)** using **RL-MPQ**
+(Reinforcement Learning Mixed-Precision Quantization): per-layer bit-width policies
+trained with PPO, validated on WikiText-2 perplexity.
 
-This repository groups **five RL-MPQ quantization scenarios** for
-[meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) in one place.
+This repo ships **five compression scenarios** as subfolders — from near-FP16 fidelity
+to aggressive survival mode — so you can pick the bits-vs-quality trade-off for your thesis experiments.
 
-Each scenario is a separate subfolder with fake-quantized FP16 weights, tokenizer, and `rlmpq_policy.json`.
+| | |
+|---|---|
+| **Base model** | [meta-llama/Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) |
+| **Method** | RL-MPQ (PPO per-layer bit policy) |
+| **Format** | Fake-quant FP16 weights + `rlmpq_policy.json` |
+| **Recommended start** | `subfolder="Balanced"` |
 
 ## Scenarios
 
@@ -25,41 +41,46 @@ Each scenario is a separate subfolder with fake-quantized FP16 weights, tokenize
 - [`Aggressive/`](./Aggressive/) — load with `subfolder="Aggressive"`
 - [`Extreme_Survival/`](./Extreme_Survival/) — load with `subfolder="Extreme_Survival"`
 
-## Summary table
+## Results (WikiText-2)
 
-| Scenario | Avg bits | Compression | WikiText-2 PPL |
-|----------|----------|-------------|----------------|
+| Scenario | Avg bits | Compression vs FP16 | Perplexity |
+|----------|----------|---------------------|------------|
 | [High_Fidelity](./High_Fidelity/) | 6.5 | 2.4615x | 4.9808 |
 | [Conservative](./Conservative/) | 5.125 | 3.122x | 5.0276 |
 | [Balanced](./Balanced/) | 4.375 | 3.6571x | 5.0437 |
 | [Aggressive](./Aggressive/) | 3.5938 | 4.4522x | 5.2614 |
 | [Extreme_Survival](./Extreme_Survival/) | 2.9688 | 5.3895x | 10.9577 |
 
-## Quick load (example: Balanced)
+## Usage
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 repo = "AvoCahDoe/llama-2-7b-rlmpq"
-scenario = "Balanced"  # or High_Fidelity, Conservative, Aggressive, Extreme_Survival
+scenario = "Balanced"  # High_Fidelity | Conservative | Aggressive | Extreme_Survival
 
 model = AutoModelForCausalLM.from_pretrained(repo, subfolder=scenario, torch_dtype="float16")
 tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=scenario)
 ```
 
-## Method
+> **Important:** Always pass `subfolder=<scenario>`. Root `config.json` describes the
+> collection; weights and tokenizer live inside each scenario folder.
 
-PPO-trained per-layer bit-width policies (Phase 3), validated with env replay + WikiText-2 PPL (Phase 4-new).
-Weights are fake-quantized in FP16 storage (not GPTQ/AWQ packed format).
+## Method (thesis summary)
 
-## Download stats
-
-Hub download counts increment when `config.json` is fetched (root index or scenario subfolder).
-Use `from_pretrained(repo, subfolder="<scenario>")` so the scenario `config.json` is requested.
-The root `config.json` is a collection index only — always pass `subfolder` to load weights.
+1. **Phase 3** — PPO agent selects per-layer bit widths under scenario-specific reward targets.
+2. **Phase 4** — Policies replayed on real weights; WikiText-2 PPL measures quality retention.
+3. **Export** — Fake-quantized FP16 checkpoints (compatible with Hugging Face Transformers).
 
 ## Citation
 
-RL-NMP-Model-Quantasation — RL-MPQ thesis framework.
+```bibtex
+@misc{rlmpq2026,
+  title  = {RL-MPQ: Reinforcement Learning Mixed-Precision Quantization},
+  author = {AvoCahDoe},
+  year   = {2026},
+  url    = {https://huggingface.co/AvoCahDoe/llama-2-7b-rlmpq}
+}
+```
 
-Generated: 2026-06-11T15:48:16.624893
+Part of the **RL-NMP-Model-Quantasation** thesis framework.