shellminator-270m-bash-distilled

shellminator-270m — a tiny natural-language → bash command model

shellminator is a terminal-native bash command assistant: type sm "copy jpgs to backup" and a 270M model suggests a single bash command; press Enter to run it in your shell, e to edit, r to refine, c to cancel. This repo is the model behind it — a Gemma-3-270M fine-tuned to translate a short natural-language request into one bash command.

Status: work-in-progress (pre-final checkpoint). The current weights were trained on the 30K qwen-distilled dataset. A combined dataset (emirkaan 6K + 30K qwen ≈ 36K) is the intended next checkpoint — see Training data and Evaluation.

Intended use

Suggesting a single bash command from a short natural-language request, with a human in the loop:

$ sm "kill the process listening on port 8080"
> kill -9 $(lsof -t -i :8080)
  [Enter] run  [e] edit  [r] refine  [c] cancel

The model only suggests — the sm UI always shows the command for review before it runs. It is not an autonomous agent.

How to use

For the sm tool the model is served as a GGUF via llama.cpp (a pre-quantized ~250MB Q4_K_M GGUF is shipped; end users never need torch). With transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer
tok  = AutoTokenizer.from_pretrained("ISB369/shellminator-270m-bash-distilled")
model = AutoModelForCausalLM.from_pretrained("ISB369/shellminator-270m-bash-distilled")
SYS = ("You are a helpful assistant that translates natural language to bash commands.\n"
       "Context: cwd=/home/user, system=Linux x86_64, shell=bash.\n"
       "Reply with a single bash command only. No explanation, no markdown fences.")
msgs = [{"role":"system","content":SYS},
        {"role":"user","content":"Generate single Bash command: list the 10 biggest files in cwd"}]
out = model.generate(**tok.apply_chat_template(msgs, tokenize=True, return_tensors="pt"),
                     do_sample=False, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True).split("assistant\n")[-1].strip())

Use greedy decoding (do_sample=False / temperature=0) — the task is deterministic; sampling hurts it.

Training

• Base: micrictor/gemma-3-270m-it-ft-bash (Gemma-3 270M, instruction-tuned + bash fine-tune).
• Method: full supervised fine-tune with TRL SFTTrainer; fp16-AMP over fp32; 3 epochs; per-epoch held-out eval with load_best_model_at_end (ships the best-generalizing epoch, not the most-trained); effective batch 16; lr 1e-4, cosine. Trained on a free Google Colab T4.
• Orphan-token fix: the base ships an orphan <image_soft_token> (id 262144) with no embedding row; training calls resize_token_embeddings(len(tokenizer)) so it gets a real trained embedding — otherwise the GGUF convert fails on the embedding-contract violation (max(token_id) < vocab_size).

Training data

The shellminator distillation pipeline: a strong cloud teacher (qwen3.5:397b-cloud via ollama /api/chat, think:false) generates varied natural-language requests, labels each with a single bash command, then filters with bash -n (syntax) + an LLM judge (correctness) and dedupes by (nl, cmd). 22 dev-tool categories: file ops, text processing, processes, networking, git (basic + advanced), docker, kubernetes, build tools, package managers, systemd, tmux, editing, cloud CLI, monitoring, permissions, disk, archives, ssh, system info, pipes/xargs, scheduling.

Datasets:

• ISB369/shellminator-bash-dataset — the 30K qwen-distilled (parts 000+001).
• ISB369/shellminator-bash-clean — 10K combined (emirkaan 6K + qwen 4K).
• ISB369/shellminator-bash-combined — 36K combined (emirkaan 6K + 30K qwen, deduped, bash -n-filtered). Intended for the final retrain.

The current checkpoint was trained on the 30K qwen alone (no emirkaan). The combined-36K retrain adds emirkaan's "echo the user's literal, specific real command" style back — see Evaluation.

Evaluation

25 held-out prompts across all 22 categories, judged by 5x majority vote of the cloud teacher (qwen3.5:397b-cloud, greedy) + bash -n validity. (Single-vote judging was too noisy; majority smooths the flip-flopping.)

What the 30K learned (win): kubectl get pods -n production (previously a ps pipeline — the kubernetes category was absorbed); unzip backup.zip -d restore.

Regressions vs the 10K-clean model: the 30K qwen data leans generic/placeholder, so the model now emits e.g. ssh-copy-id user@remote_host instead of echoing the user's literal 192.168.1.10, and is sloppier on precise tasks. This is the data style > scale lesson — more coverage, but the style shift cost precision. The combined-36K retrain is designed to fix this (emirkaan restores literal-echoing while keeping the dev-tool coverage).

Capacity ceiling (not a data issue): the model sometimes mangles complex syntax (e.g. awk with nested quotes, an unbalanced paren). The dataset has zero invalid-syntax labels, so these are 270M generation limits, not bad data. A bigger model or constrained decoding (llama.cpp GBNF grammars) is the fix there.

Limitations

• 270M capacity — strong on common single-line commands; mangles complex multi-arg/nested syntax; no multi-step reasoning.
• Can be wrong or destructive — may suggest an incorrect or dangerous command (wrong flags, rm/kill with wrong targets). Always review before running.
• English requests, Linux/bash, x86_64 only.
• Literal-precision is the current (30K) checkpoint's weak spot (placeholders); the combined retrain targets this.

Safety / ethics

The model suggests shell commands that can modify or delete data. The sm tool is built so a command is never run without a human pressing Enter (shown for review; e to edit, c to cancel). Do not wire this model into an autonomous executor. Treat every suggestion as untrusted until you have read it.

Reproduce

SFTTrainer on the combined dataset (HF_DATASET_REPO=ISB369/shellminator-bash-combined), Colab T4, the config above. Eval: train/eval.py (25 prompts, 5x majority judge). Generation pipeline: scripts/generate_dataset.py (teacher → NL → bash → bash -n + judge → dedup; resumable; HF upload). Merge: scripts/merge_combined.py.

---

Built as a distillation exercise: a big cloud model's behavior becomes a tiny local model's training data, so a ~250MB model can suggest a bash command in under a second on CPU — privately, with no GPU.