Hugging Face's logo
VIDRAFT_LAB's picture
In a Training Loop πŸ”„
18 11 95

VIDRAFT_LAB

SeaWolf-AI
LOYOLABIZ's profile picture John6666's profile picture huggingworld's profile picture
Β·

AI & ML interests

None yet

Recent Activity

updated a collection about 2 hours ago
DARWIN-Family
liked a model about 2 hours ago
FINAL-Bench/Darwin-28B-Opus
reacted to theirpost with πŸ”₯ about 6 hours ago
🧬 Introducing Darwin-9B-NEG β€” the first model with Native Entropy Gating (NEG) πŸ”— Try it now: https://huggingface.co/FINAL-Bench/Darwin-9B-NEG We're thrilled to release Darwin-9B-NEG, a 9B-parameter reasoning model that embeds an architecturally-internalised sense of self-confidence directly into the transformer β€” our proprietary Native Entropy Gating (NEG) technology. πŸ“Š GPQA Diamond (198 PhD-level questions): β–Έ Baseline Darwin-9B (no NEG) β†’ 51.01 % β–Έ Pure NEG (greedy Β· 1Γ— cost) β†’ 63.64 % πŸ”₯ +12.63 %p β–Έ + Permutation (4Γ— cost) β†’ 76.26 % β–Έ + Ensemble Refinement (~20Γ—) β†’ 84.34 % πŸ† With only 9 billion parameters and 1Γ— inference cost, Pure NEG jumps +12.63 %p over the same model without NEG. Going all-in with ensemble refinement pushes it to 84.34 % β€” surpassing the published Qwen3.5-9B leaderboard score (81.7 %) by +2.64 %p. πŸ”¬ What makes NEG different from Multi-Turn Iteration (MTI)? Classical MTI needs 3-8Γ— extra inference passes. NEG instead lives INSIDE the single decoding loop. Two tiny modules ride with the transformer: NEG-Head predicts per-token entropy from the last hidden state, and NEG-Gate conditionally restricts the top-k choice when confidence is low. The gate activates in only 4.36 % of tokens β€” essentially free at inference time. ✨ Key differentiators β€’ Architecturally internalised β€” model file *is* the feature β€’ 1Γ— inference cost (vs. 3-8Γ— for MTI) β€’ Drop-in with vLLM / SGLang / TGI / transformers β€” no extra engine β€’ +12.63 %p reasoning at zero latency overhead β€’ Single-file deployment, Apache 2.0 licensed 🧬 Lineage Qwen/Qwen3.5-9B β†’ Darwin-9B-Opus (V7 evolutionary merge) β†’ Darwin-9B-NEG (V8 + NEG training) #Darwin #NEG #NativeEntropyGating #GPQA #Reasoning #LLM #OpenSource #Apache2
View all activity

Organizations

FINAL_Bench's profile picture