ConvFill Gemma E4B LoRA (cold predictor)
Phase 0 plumbing smoke. LoRA SFT adapter for the ConvFill fast foreground predictor, trained on the mined cold-predictor data.
- base model:
google/gemma-4-E4B-it (text tower via Gemma4ForCausalLM; UNGATED)
- data: contract-shaped prompt/completion from
data/sft_v0.jsonl (stage3_assemble.py; verdict->training-role rule)
- examples: 727
- LoRA: r=16, alpha=32, dropout=0.05, targets=['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']
- recipe: lr=0.0002, epochs=3.0, warmup_ratio=0.03, completion_only_loss=True, packing=False
This repo holds the ADAPTER only — merge into the base, convert to GGUF, and serve via Ollama (see PHASE0_RUNBOOK.md).
