LFM2.5-350M-ShellAI
ShellAI Bash-command model distilled at the response level from
LiquidAI/LFM2.5-2.6B into LiquidAI/LFM2.5-350M.
This repository contains the merged BF16 training checkpoint.
The model is trained to emit exactly one command inside:
<shellai-command>command</shellai-command>
General chat anchors, assistant-only loss, LoRA, one training epoch, early stopping, and a pre-publication retention gate are used to reduce catastrophic forgetting. No generated command was executed during dataset construction or evaluation.
Distillation
The two models have different vocabularies (128K teacher versus 65,536 student), so this uses sequence-level response distillation rather than token-logit KL. Teacher candidates that fail the command envelope, primary-utility, or reference- similarity gates fall back to the verified dataset reference.
Evaluation
Held-out repository test split (300 examples, BF16 Transformers):
| Model | Exact | Utility match | Token F1 | Valid envelope |
|---|---|---|---|---|
| Base 350M | 5.7% | 23.0% | 0.225 | 100.0% |
| Distilled 350M | 4.0% | 39.7% | 0.305 | 100.0% |
Chat retention used 50 non-shell prompts: shell-envelope leakage remained 0.0%; anchor similarity retained 84.9% of baseline.
Q8_0 llama.cpp CPU test (same 50-example subset):
| Threads | Token F1 | Utility match | Median latency | Decode |
|---|---|---|---|---|
| 1 | 0.268 | 36.0% | 1278 ms | 16.6 tok/s |
| 2 | 0.268 | 36.0% | 641 ms | 32.5 tok/s |
License
This is a modified derivative of Liquid AI's LFM2.5 weights and is distributed
under the included LFM Open License v1.0. See NOTICE for modification details.