LFM2.5-350M-ShellAI-v2-GGUF

LFM2.5-350M ShellAI v2 — two-stage Q8_0

This release translates a natural-language shell request through a validated typed intermediate representation before producing Bash:

1. LFM2.5-350M-ShellAI-v2-IR-Q8_0.gguf: natural language to ShellIR-v1.
2. LFM2.5-350M-ShellAI-v2-Command-Q8_0.gguf: ShellIR-v1 to one Bash command.

Both files are required for the learned two-stage pipeline. A deterministic ShellIR compiler can replace the second model when minimum memory or latency is more important than retaining the learned lowering stage. The untouched LiquidAI conversational base remains usable without either task adapter.

These are ordinary post-training Q8_0 GGUF files. They are not Liquid AI QAD checkpoints and are not labeled as QAD.

Evaluation

The sealed suite contains 20 compositional NL-to-Bash cases. No generated command was executed.

The one Q8_0-only regression was a permission-mask choice in the world-writable-files case. Thread count did not change model output.

Measurements used llama.cpp b10516 with CPU-only inference on the development machine. See hard_suite_two_stage.json and q8_0_two_stage_gguf.json for per-case results.

Training and selection

The Stage-1 release is an exact LoRA-delta blend of 87.5% safety-focused and 12.5% balanced checkpoints. It improved strict suite accuracy from 70% to 80% while retaining 100% risk/effects-label accuracy. Stage 2 was 100% exact when given gold ShellIR.

Treat model output as untrusted. Validate ShellIR, enforce the risk/effects policy, show destructive commands to the user, and never execute a generated command without explicit authorization.

License

This is a modified derivative of LiquidAI/LFM2.5-350M, distributed under the included LFM Open License v1.0. See NOTICE for the modification statement.