RL-MPQ — Llama 3 8B - 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 — Llama 3 8B

updated Jun 11

5 RL-MPQ fake-quant Llama 3 8B scenarios: High Fidelity through Extreme Survival. Thesis checkpoints.

AvoCahDoe/llama-3-8b-rlmpq-aggressive

Text Generation • 8B • Updated Jun 11 • 4

Note Aggressive: 3.7188 bits/weight, 4.3025x vs FP16, WikiText-2 PPL 6.6793. RL-MPQ PPO fake-quant of meta-llama/Meta-Llama-3-8B.
AvoCahDoe/llama-3-8b-rlmpq-balanced

Text Generation • 8B • Updated Jun 11 • 6

Note Balanced: 4.5 bits/weight, 3.5556x vs FP16, WikiText-2 PPL 6.0199. RL-MPQ PPO fake-quant of meta-llama/Meta-Llama-3-8B.
AvoCahDoe/llama-3-8b-rlmpq-conservative

Text Generation • 8B • Updated Jun 11 • 6

Note Conservative: 5.25 bits/weight, 3.0476x vs FP16, WikiText-2 PPL 5.9652. RL-MPQ PPO fake-quant of meta-llama/Meta-Llama-3-8B.
AvoCahDoe/llama-3-8b-rlmpq-extreme-survival

Text Generation • 8B • Updated Jun 11 • 3

Note Extreme Survival: 2.875 bits/weight, 5.5652x vs FP16, WikiText-2 PPL 32.393. RL-MPQ PPO fake-quant of meta-llama/Meta-Llama-3-8B.
AvoCahDoe/llama-3-8b-rlmpq-high-fidelity

Text Generation • 8B • Updated Jun 11 • 7

Note High Fidelity: 6.875 bits/weight, 2.3273x vs FP16, WikiText-2 PPL 5.8133. RL-MPQ PPO fake-quant of meta-llama/Meta-Llama-3-8B.