Model Card: GenomeOcean-100M (FP8)
Generated: 2026-05-09T18:03:44-0700
Architecture
| Parameter | Value |
|---|
| Architecture | MistralForCausalLM |
| Model Type | mistral |
| Vocab Size | 4096 |
| Hidden Size | 768 |
| Num Hidden Layers | 12 |
| Num Attention Heads | 8 |
| Intermediate Size | 3072 |
| Max Position Embeddings | 32768 |
| RoPE Theta | 1000000.0 |
Quantization Method
- Format: FP8 (E4M3) per-channel weight-only quantization
- Scale DType: float32 per-channel scales
- Method: Post-training quantization (PTQ) with per-channel E4M3 weights
Perplexity Results
| Metric | Value |
|---|
| Original PPL (BF16) | 40804.858 |
| Quantized PPL (FP8) | 41000.839 |
| PPL Difference | 195.981 |
| PPL Difference (%) | 0.48% |
Quality Assessment: Excellent - negligible quality loss
Weight Fidelity
| Metric | Value |
|---|
| Mean Cosine Similarity | 0.999581 |
| Min Cosine Similarity | 0.999509 |
| Mean Relative L2 Error | 0.026446 |
| Max Relative L2 Error | 0.026671 |
| Layers Compared | 85 |
Compression
| Metric | Value |
|---|
| Original Size | 0.2394 GB |
| Quantized Size | 0.1171 GB |
| Compression Ratio | 48.91% |
| Space Saved | 0.12 GB |
Summary
The GenomeOcean-100M model was quantized from BF16 to FP8.
Perplexity changed by 0.48% (original: 40804.858, quantized: 41000.839).
Mean weight cosine similarity is 0.9996.
Compression ratio is 48.91% (saved 0.12 GB).
