RL-MPQ — Gemma 2 9B - a AvoCahDoe Collection

AvoCahDoe 's Collections

Invoice Layout Extraction

RL-MPQ VLM — Qwen2-VL-7B

RL-MPQ VLM — LLaVA-Next Mistral-7B

RL-MPQ VLM — LLaVA-1.5-7B

RL-MPQ VLM — LLaVA-1.5-13B

RL-MPQ LLaVA-1.5-13B VLM Evaluation

RL-MPQ — Qwen 2.5 7B

RL-MPQ — Llama 2 7B

RL-MPQ — Llama 3 8B

RL-MPQ — Llama 2 13B

RL-MPQ — Mistral 7B

RL-MPQ — Gemma 2 9B

RL-MPQ — Llama 3.1 8B

RL-MPQ — JetMoE 8B

RL-MPQ — Gemma 2 9B

updated Jun 11

5 RL-MPQ fake-quant Gemma 2 9B scenarios. Per-layer bit-width policies from Phase 3 PPO.

AvoCahDoe/gemma-2-9b-rlmpq-aggressive

Text Generation • 9B • Updated Jun 11 • 23

Note Aggressive: 3.6667 bits/weight, 4.3636x vs FP16, WikiText-2 PPL 162.8437. RL-MPQ PPO fake-quant of google/gemma-2-9b.
AvoCahDoe/gemma-2-9b-rlmpq-balanced

Text Generation • 9B • Updated Jun 11 • 26

Note Balanced: 4.2857 bits/weight, 3.7333x vs FP16, WikiText-2 PPL 127.0798. RL-MPQ PPO fake-quant of google/gemma-2-9b.
AvoCahDoe/gemma-2-9b-rlmpq-conservative

Text Generation • 9B • Updated Jun 11 • 19

Note Conservative: 5.1429 bits/weight, 3.1111x vs FP16, WikiText-2 PPL 116.5244. RL-MPQ PPO fake-quant of google/gemma-2-9b.
AvoCahDoe/gemma-2-9b-rlmpq-extreme-survival

Text Generation • 9B • Updated Jun 11 • 16

Note Extreme Survival: 2.7857 bits/weight, 5.7436x vs FP16, WikiText-2 PPL 424.7991. RL-MPQ PPO fake-quant of google/gemma-2-9b.
AvoCahDoe/gemma-2-9b-rlmpq-high-fidelity

Text Generation • 9B • Updated Jun 11 • 21

Note High Fidelity: 7.0476 bits/weight, 2.2703x vs FP16, WikiText-2 PPL 104.8098. RL-MPQ PPO fake-quant of google/gemma-2-9b.