🧠 mohommed/LFM3-1.7B-Frontier-SmartAgentv4 (SDT-v2 • Assistant-Masked • ShortConv-Unfrozen • SimPO • 512K)

LFM3-1.7B-Frontier-SmartAgentv4 is trained via SDT-v2 (Smart Deep Training v2) across Dual-T4 GPUs:

1. 100% Assistant-Turn Loss Masking + Focal Surprise Gate $(1 - p_t)^{1.5}$ + Pivotal Token Boost ($1.85\times$).
2. Unfrozen Native shortconv.conv 1D Convolutions & QK-Norms alongside Rank-Stabilized LoRA (rsLoRA r=64, alpha=128).
3. Reference-Free SimPO Contrastive Margin Pass (Winner $y^+$ vs. Trap/Slop Loser $y^-$).
4. Wise-FT Task-Vector Anchor Blending (0.88 / 0.12) + Post-Merge Triple-Subspace SVD Uncensoring.
5. 512K Fast ABF-RoPE + llama-imatrix Calibrated GGUFs: Hosted in mohommed/LFM3-1.7B-Frontier-SmartAgentv4-GGUF.