Downloads · 30 days
0
DDDDD-433/speaksh-weighted-prompt-v1-200
speaksh-weighted-prompt-v1-200 is a text generation model from DDDDD-433. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as mit.
LoRA adapter for speaksh, a local-first natural-language-to-shell command CLI for Unix-like systems.
Downloads · 30 days
0
Access
Public
Updated Jun 22, 2026
Repo size
15 MB
Likes
0
Public
Click a slice to open those files.
.safetensors15 MB · 100%
From the Hugging Face model README
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.
mlx-community/MiniCPM5-1B-OptiQ-4bit
Use this adapter with the speaksh CLI for local command suggestion
experiments. It is intended for reviewed, interactive use, not unattended shell
automation.
Example:
speaksh \
--adapter-path adapters/speaksh-weighted-prompt-v1-200 \
--dry-run "find files over 100mb"
The training pipeline uses permissive public NL-to-shell datasets:
westenfelder/NL2SH-ALFA - MITemirkaanozdemr/bash_command_data_6K - Apache-2.0AryaYT/nl2shell-training-v3 - Apache-2.0It also includes a small MIT-licensed curated set for supported speaksh
command families:
Filtering removes dangerous commands, multiline commands, empty examples, and
benchmark-specific /testbed paths before training.
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
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.