SS-Talk-2-Bash-GGUF

Model Card: SS-Talk-2-Bash (GGUF Version)

This is the GGUF quantized version of saadxsalman/SS-Talk-2-Bash. This format is specifically designed for high-performance inference on CPUs and local hardware using tools like llama.cpp, LM Studio, or Ollama.

---

1. Model Description

• Developed by: saadxsalman
• Model type: Causal Language Model (LFM)
• Base Model: LiquidAI/LFM2.5-350M
• Format: GGUF (optimized for local deployment)
• Training Goal: Deterministic Natural Language to Bash command translation.

---

2. Training Strategy: "The Hard-Coding Engine"

This model utilizes a Strict Hard-Coding training method. It was designed to function as a tool rather than a chatty assistant:

• Zero Chatter: The model is trained to jump directly from the [CL] token to the command syntax.
• Label Masking: Loss was only calculated on the Bash output, making the model highly specialized for structural accuracy over conversational fluidity.
• Deterministic: Ideally used with temperature: 0.0 for consistent, repeatable results.

---

3. Intended Use & Prompt Format

The model is extremely sensitive to formatting. You must use the specific trigger tokens to get accurate results:

Correct Prompt Template: [NL] {Your natural language request here} [CL]

Example: [NL] find all files larger than 100MB in the current directory [CL]

---

4. How to Use with GGUF

via llama.cpp (CLI)

./llama-cli -m ss-talk-2-bash-q8_0.gguf \
  -p "[NL] list all files in long format [CL]" \
  --temp 0 \
  -n 64 \
  --stop "[END]"

via llama-cpp-python

from llama_cpp import Llama

llm = Llama(model_path="ss-talk-2-bash-q4_k_m.gguf")

prompt = "[NL] list all docker containers [CL]"
output = llm(prompt, max_tokens=64, temperature=0.0, stop=["[END]"])

print(output['choices'][0]['text'])

---

5. Quantization Comparison

---

6. Limitations

• Logic Only: This model has "forgotten" how to converse. It will not answer general knowledge questions.
• Formatting Sensitive: If the [NL] or [CL] tokens are omitted, performance will degrade significantly.
• Scope: Optimized for standard Linux Bash commands; complex scripting logic outside the training data may require manual review.

---

7. Training Metadata

• Dataset: emirkaanozdemr/bash_command_data_6K
• Method: PEFT (LoRA)
• Hyperparameters: $1 \times 10^{-4}$ Learning Rate, R64 LoRA Alpha 128.
• License: Apache 2.0