Access PetInst-LLM under the Gemma Terms
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a modified Gemma model derivative. By accessing it, you agree to the Gemma Terms of Use and incorporated Prohibited Use Policy linked in this repository.
Log in or Sign Up to review the conditions and access this model content.
PetInst-LLM 1B v6.18 — MLX Q5 + LoRA
PetInst-LLM 1B is an independently modified research derivative of Google's google/gemma-3-1b-it, specialized for one silent, schema-constrained virtual-pet function decision.
This project is not affiliated with, sponsored by, or endorsed by Google. Google DeepMind created the upstream Gemma model; the PetInst-LLM project created and evaluated these modifications.
Packaging
This repository preserves the exact selected research runtime:
base/: recovered MLX affine 5-bit base from the PetInst research lineageadapter/: selected v6.18 MLX LoRA adapterruntime.json: input, context, serialization, and decoding contractevaluation/: aggregate evidence without private transcripts or raw audio
This is not a GGUF and is not directly loadable by llama.rn. Earlier merged/requantized 1B artifacts regressed substantially, so they are intentionally not published as equivalent deployment models.
Evaluation
On fresh 220-row Control D with v1.3 descriptions and constrained selection among the declared complete calls:
| Metric | Result |
|---|---|
| Exact/tool accuracy | 97.27% |
| Memory-policy accuracy | 95% |
| Memory read/write | 100% / 100% |
| Ineligible memory-write false positives | 0% |
| Negative-expression accuracy | 90% |
| State-precedence accuracy | 100% |
| Safety accuracy | 100% |
| Mean / P95 / max latency | 1.203 / 1.442 / 2.140 s |
| Recorded peak memory | 1.632 GB |
The model passed 29 of 35 disclosed research checks. Order consistency and latency remained below the project gate. Results are desktop MLX evidence, not physical-device or airplane-mode proof.
Intended use
- Research comparison against the compact PetInst-LLM 270M intent policy
- MLX desktop evaluation of high-level virtual-pet tool selection
- Continued study of memory consent, state precedence, safety, negative expression, and tool-order robustness
Limitations
- Research-only and release-unapproved.
- Not compatible with the current Expo
llama.rnpath in this form. - Requires constrained decoding over three to five complete declared calls.
- The authoritative pet state, wallet, inventory, purchases, memory retention, and policy enforcement must remain deterministic host responsibilities.
- Only English synthetic research data was evaluated.
Terms and modification notice
Access and use are subject to the Gemma Terms of Use and the incorporated Gemma Prohibited Use Policy. See NOTICE and MODIFICATIONS.md.
Quantized