---
language:
- en
license: apache-2.0
library_name: peft
pipeline_tag: text-generation
base_model: LiquidAI/LFM2.5-350M
tags:
- bash
- terminal
- devops
- linux
- hardcoded
- lfm
datasets:
- emirkaanozdemr/bash_command_data_6K
model_creator: saadxsalman
widget:
- text: "[NL] find all files larger than 100MB in the current directory [CL]"
  example_title: "Find Large Files"
- text: "[NL] list all files in long format including hidden ones [CL]"
  example_title: "List All Files"
inference:
  parameters:
    do_sample: false
    temperature: 0.0
    max_new_tokens: 64
---

# SS-Talk-2-Bash (LFM-350M-Hardcoded)

This model is a fine-tuned version of **LiquidAI/LFM2.5-350M** designed specifically for deterministic natural language to Bash command translation. It uses a **Strict Hard-Coding** training method to minimize linguistic "chatter" and maximize structural accuracy.

---

## 1. Model Description
* **Developed by:** saadxsalman
* **Model type:** Causal Language Model (LFM)
* **Language(s):** English (Input) to Bash (Output)
* **License:** Apache 2.0
* **Finetuned from model:** LiquidAI/LFM2.5-350M

---

## 2. Training Strategy: "The Hard-Coding Engine"
Unlike standard instruction-tuned models that learn to be "helpful assistants," this model was trained using a **Masking Collator** strategy:

* **Label Masking:** All natural language tokens (the prompt) are masked during training ($loss = -100$). The model only calculates loss on the Bash command itself.
* **Zero Chatter:** The model does not learn to say "Sure, here is your command." It is trained to jump directly from the `[CL]` token to the syntax.
* **Greedy Decoding:** The `generation_config.json` is locked to `do_sample: False` and `temperature: 0.0` to ensure the same input always produces the same output.

---

## 3. Training Data
The model was fine-tuned on the `emirkaanozdemr/bash_command_data_6K` dataset. The data was restructured into a rigid non-linguistic format:
`[NL] {Natural Language Prompt} [CL] {Bash Command} [END]`

---

## 4. Intended Use & Prompting
To get the best results, you **must** use the specific trigger tokens used during training.

**Correct Prompt Format:**
`[NL] find all files larger than 100MB in the current directory [CL]`

---

## 5. How to Use (Inference)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saadxsalman/SS-Talk-2-Bash"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "[NL] list all files in long format [CL]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## 6. Limitations and Biases
* **Logic Only:** This model has "forgotten" how to converse. It will not answer general questions or write Python code.
* **Bash Specific:** It is optimized for standard Linux Bash commands. It may struggle with complex shell scripting logic if not represented in the 6K training samples.
* **Formatting Sensitive:** If the `[NL]` or `[CL]` tokens are omitted, the model performance will degrade significantly.

---

## 7. Training Hyperparameters
| Parameter | Value |
| :--- | :--- |
| **Learning Rate** | $1 \times 10^{-4}$ |
| **Optimizer** | Paged AdamW 8-bit |
| **LoRA R** | 64 |
| **LoRA Alpha** | 128 |
| **Batch Size** | 16 (4 per device $\times$ 4 grad accum) |
| **Precision** | Mixed Precision (FP16) |

```