Experimental fine-tune of google/gemma-4-E2B (eliza-1-2b tier) specialized on
the elizaOS agent harness. Trained on trajectories harvested from gpt-5.5
(via the Codex subscription) running the elizaOS scenario / benchmark / e2e
suites, so the model leans into our native tool-calling and JSON-decision shapes.
Base: google/gemma-4-E2B
Method: APOLLO SFT, 1 epoch, on a 2,774-row combined corpus (55% gpt-5.5
scenario trajectories + 45% base eliza-1 SFT corpus). Training data:
elizaos/eliza-1-training-data (converted/gpt55-scenarios/).
eliza-1-2b-gpt55.q4_k_m.gguf — 3.2 GB, mobile/desktop-deployable. Verified
generating coherent text at 157 tok/s on Apple M4 Max (Metal, llama.cpp).
eliza-1-2b-gpt55.f16.gguf — 8.6 GB, full-precision source for re-quantizing.
Honest status
Experimental. Light (1-epoch) specialization pass. On a small local benchmark,
native-tool-call / JSON structure adherence moved 50% → 62.5% vs the base
(directional, tiny sample — a full-split GPU benchmark is future work). Not a
gated official eliza-1 release; use as an experimental checkpoint.
llama-cli -m eliza-1-2b-gpt55.q4_k_m.gguf -p "List three things a helpful assistant does:" -n 64
