Views
No views yet
| Parameter | Value |
|---|---|
| start_layer_index | 20 |
| end_layer_index | 36 |
| preserve_good_behavior_weight | 0.5805 |
| steer_bad_behavior_weight | 0.0023 |
| overcorrect_relative_weight | 0.7764 |
| neighbor_count | 13 |
| Metric | Heretic | Original QAT |
|---|---|---|
| KL Divergence | 0.0065 | 0 (by definition) |
| Refusals | 5/100 | 98/100 |
| Base Model | google/gemma-4-E4B-it |
| Parameters | 4.5B effective (8B with PLE embeddings) |
| Architecture | Dense with Per-Layer Embeddings (PLE) + multimodal |
| Layers | 42 |
| Hidden Size | 2,560 |
| Context Length | 128K tokens |
| Vocabulary | 262K, 140+ languages |
| Modalities | Text + Image + Audio + Video |
| Quantization | Q4_0 (matching Unsloth UD-Q4_K_XL layout), b9553 llama-quantize |
| Format | Q4_0 (QAT weights optimized for this precision) |
| File Size | 4 GB |
| Effective BPW | 4.54 (all weights Q4_0, norms F32) |
| Tool | llama-quantize (b9553, CUDA 13.3) |
| QAT Advantage | Q4_0 + QAT weights achieves 98.5% Top-1 vs 90.9% naive Q4_0 (+7.6%) |
| General | temp=1.0, top_p=0.95, top_k=64 |
| Coding | temp=0.6, top_p=0.95, top_k=64 |
--jinja flag with llama.cpp. Disable thinking: --chat-template-kwargs '{"enable_thinking":false}'.