--- license: mit base_model: mlx-community/MiniCPM5-1B-OptiQ-4bit library_name: mlx tags: - text-generation - shell - cli - mlx - lora - natural-language-to-code --- # speaksh-weighted-prompt-v1-200 LoRA adapter for `speaksh`, a local-first natural-language-to-shell command CLI for Unix-like systems. The adapter targets short command-generation requests for file search, disk usage, process lookup, archive creation, package installation, and Git status. The `speaksh` runtime applies deterministic safety checks and narrow command canonicalization after generation. ## Base Model ```text mlx-community/MiniCPM5-1B-OptiQ-4bit ``` ## Intended Use Use this adapter with the `speaksh` CLI for local command suggestion experiments. It is intended for reviewed, interactive use, not unattended shell automation. Example: ```bash speaksh \ --adapter-path adapters/speaksh-weighted-prompt-v1-200 \ --dry-run "find files over 100mb" ``` ## Training Data The training pipeline uses permissive public NL-to-shell datasets: - `westenfelder/NL2SH-ALFA` - MIT - `emirkaanozdemr/bash_command_data_6K` - Apache-2.0 - `AryaYT/nl2shell-training-v3` - Apache-2.0 It also includes a small MIT-licensed curated set for supported `speaksh` command families: - note-aware package manager selection - size queries - file search - port and process lookup - compression commands Filtering removes dangerous commands, multiline commands, empty examples, and benchmark-specific `/testbed` paths before training. ## Training Command ```bash mlx_lm.lora \ --model mlx-community/MiniCPM5-1B-OptiQ-4bit \ --train \ --data data/processed/speaksh_public_v1/mlx \ --adapter-path adapters/speaksh-weighted-prompt-v1-200 \ --iters 200 \ --batch-size 2 \ --learning-rate 2e-5 \ --max-seq-length 512 \ --steps-per-report 20 \ --steps-per-eval 50 \ --val-batches 25 \ --save-every 100 \ --seed 42 ``` ## Results ```text Final validation loss: 0.279 Test loss: 0.241 Test perplexity: 1.273 Baseline speaksh eval: 37/37 Held-out speaksh eval: 40/40 External public fallback eval: 32/32 Raw adapter baseline score before command canonicalization: 30/37 ``` The `speaksh` eval scores include runtime safety classification and narrow command canonicalization. ## Safety And Limitations - This adapter is not a sandbox. - Commands should be reviewed before execution. - The runtime blocks commands classified as dangerous unless explicitly overridden. - The benchmark is exact-match oriented and does not cover every valid shell formulation. - Windows and PowerShell are not supported. ## Links - GitHub: https://github.com/DDDD-433/speaksh
