Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
base_model: deepseek-ai/DeepSeek-V4-Flash-0731
|
| 4 |
+
library_name: transformers
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# Smaug-Flash
|
| 8 |
+
|
| 9 |
+
Abacus.AI finetune of DeepSeek-V4-Flash-0731 (full merged weights, checkpoint-native
|
| 10 |
+
FP8: e4m3 + e8m0 128x128-block scales on the finetuned attention matrices, experts in
|
| 11 |
+
the release packed-FP4 layout — loads exactly like the official release).
|
| 12 |
+
|
| 13 |
+
Training: agentic teacher-agreement SFT + product loop-escape KTO + instruction-following
|
| 14 |
+
repair SFT, merged as full deltas (no rank truncation).
|
| 15 |
+
|
| 16 |
+
LiveBench (2026 fresh sets, clean serving): overall-7 **77.61** vs base 74.17 —
|
| 17 |
+
math 91.3 / reasoning 86.2 / data_analysis 79.1 / language 79.2 / coding 77.1 /
|
| 18 |
+
instruction_following 69.3 / agentic_coding 61.1 (base 46.8).
|
| 19 |
+
|
| 20 |
+
Serving: SGLang v0.5.18, tp4, dp-attention, --moe-runner-backend flashinfer_mxfp4,
|
| 21 |
+
fp8 KV, 1M context. Recommended reasoning effort: max.
|