Model Card: GenomeOcean-500M-v1.2 (GGUF_Q8_0)
Generated: 2026-05-09T20:11:04-0700
Architecture
| Parameter | Value |
|---|
| Architecture | MistralForCausalLM |
| Model Type | mistral |
| Vocab Size | 4096 |
| Hidden Size | 1536 |
| Num Hidden Layers | 14 |
| Num Attention Heads | 8 |
| Intermediate Size | 6144 |
| Max Position Embeddings | 32768 |
| RoPE Theta | 1000000.0 |
Quantization Method
- Format: GGUF Q8_0 (8-bit integer quantization)
- Method: Post-training quantization via llama.cpp
Perplexity Results
| Metric | Value |
|---|
| Original PPL (BF16) | 41887.0458 |
| Quantized PPL (GGUF_Q8_0) | 41846.9413 |
| PPL Difference | -40.1045 |
| PPL Difference (%) | -0.1% |
Quality Assessment: Excellent - negligible quality loss
Weight Fidelity
| Metric | Value |
|---|
| Mean Cosine Similarity | 1.000160 |
| Min Cosine Similarity | 0.999963 |
| Mean Relative L2 Error | 0.004376 |
| Max Relative L2 Error | 0.011212 |
| Layers Compared | 129 |
Compression
| Metric | Value |
|---|
| Original Size | 1.0826 GB |
| Quantized Size | 0.5812 GB |
| Compression Ratio | 53.68% |
| Space Saved | 0.50 GB |
Summary
The GenomeOcean-500M-v1.2 model was quantized from BF16 to GGUF_Q8_0.
Perplexity changed by -0.1% (original: 41887.0458, quantized: 41846.9413).
Mean weight cosine similarity is 1.0002.
Compression ratio is 53.68% (saved 0.50 GB).
