EcoSLM-WebDM (fragilov/ecoslm-web-dm)

EcoSLM-WebDM is a lightweight (~10.5M parameter) Cross-Element Pointer Transformer distilled from TypeSafe Jev (jev-latest) soft probability distributions for ultra-fast local web navigation and interaction.

Architecture & Speculative Fan-out

  • Context Encoder: 4-layer Transformer (d_model=256, nhead=4) encoding Goal + Page Title/URL + History + Visible Text
  • Element Encoder + Cross-Attention: 2-layer Element Transformer + 2-layer Cross-Attention Decoder over up to 250 DOM controls
  • Dual Speculative Heads (1 forward pass):
    • Operation Head: CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED
    • Target Pointer Head: Scores all compatible interactive DOM targets simultaneously
  • Confidence Calibration: Post-training Temperature Scaling (op_temp_scaler & target_temp_scaler) for Confidence-Gated fallback.

Measured Distillation Results

Metric Value
Parameters 10,830,085
Train / Val Samples 80 / 35
Operation Top-1 Agreement 80.00%
Target Top-1 Agreement 83.08%
Target Top-3 Agreement 100.00%
Calibrated Operation ECE 0.0964
Student CPU Median Latency 18.57 ms
Teacher Jev API Mean Latency 276.01 ms
Downloads last month
46
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support