| --- |
| license: llama3 |
| base_model: meta-llama/Meta-Llama-3-8B |
| library_name: llama.cpp |
| pipeline_tag: text-generation |
| tags: |
| - gguf |
| - rest-api-testing |
| - software-testing |
| - llama.cpp |
| - llamaresttest |
| language: |
| - en |
| --- |
| |
| # LlamaREST-IPD (GGUF) |
|
|
| **LlamaREST-IPD** is a Small Language Model (SLM) fine-tuned to predict |
| **inter-parameter dependencies** (IPDs) among REST API parameters, used by |
| [**LlamaRestTest**](https://github.com/codingsoo/LlamaRestTest) to improve automated |
| REST API test generation. |
|
|
| It is one of the two SLMs used by LlamaRestTest; the companion model, |
| [**LlamaREST-EX**](https://huggingface.co/codingsoo/LlamaREST-EX-GGUF), generates |
| realistic example values for parameters. |
|
|
| ## Models |
|
|
| Quantized with `llama.cpp` (Q6_K) from a Llama-3-8B base fine-tuned with QLoRA. |
| |
| | Variant | File | Size | |
| |---|---|---| |
| | 2B | [`LlamaREST-IPD-2B.gguf`](https://huggingface.co/codingsoo/LlamaREST-IPD-GGUF/resolve/main/LlamaREST-IPD-2B.gguf) | ~3.2 GB | |
| | 4B | [`LlamaREST-IPD-4B.gguf`](https://huggingface.co/codingsoo/LlamaREST-IPD-GGUF/resolve/main/LlamaREST-IPD-4B.gguf) | ~4.9 GB | |
| | 8B | [`LlamaREST-IPD-8B.gguf`](https://huggingface.co/codingsoo/LlamaREST-IPD-GGUF/resolve/main/LlamaREST-IPD-8B.gguf) | ~8.5 GB | |
| | F16 | [`Llama3-Ipd-Rest-8.0B-F16.gguf`](https://huggingface.co/codingsoo/LlamaREST-IPD-GGUF/resolve/main/Llama3-Ipd-Rest-8.0B-F16.gguf) | ~16 GB | |
| |
| ## Download |
| |
| ```bash |
| pip install -U "huggingface_hub[cli]" |
| hf download codingsoo/LlamaREST-IPD-GGUF LlamaREST-IPD-8B.gguf --local-dir . |
| ``` |
| |
| ## Usage |
| |
| Intended to be run through the [LlamaRestTest](https://github.com/codingsoo/LlamaRestTest) |
| pipeline. To run the GGUF directly with `llama.cpp`: |
| |
| ```bash |
| ./llama-cli -m LlamaREST-IPD-8B.gguf -p "your prompt" |
| ``` |
| |
| ## Training |
| |
| Fine-tuned from Llama-3-8B with QLoRA, then quantized with `llama.cpp` (Q6_K). Key hyperparameters: |
| |
| - 4-bit precision (nf4), compute dtype `float16`, no nested quantization |
| - LoRA: `r=64`, `alpha=16`, dropout `0.1` |
| - 5 epochs, batch size 4, `paged_adamw_32bit`, LR `2e-4`, weight decay `0.001`, constant schedule, warmup ratio `0.03` |
| |
| Training data (Inter-Parameter Dependency): [random1234321/REST-IPD](https://huggingface.co/datasets/random1234321/REST-IPD). |
| |
| ## License |
| |
| The training/testing code is released under the MIT License. The model weights are |
| derived from Meta Llama 3 and are therefore subject to the |
| [Meta Llama 3 Community License](https://llama.meta.com/llama3/license/). |
| |
| ## Citation |
| |
| ```bibtex |
| @article{kim2025llamaresttest, |
| title={LlamaRestTest: Effective REST API Testing with Small Language Models}, |
| author={Kim, Myeongsoo and Sinha, Saurabh and Orso, Alessandro}, |
| journal={arXiv preprint arXiv:2501.08598}, |
| year={2025} |
| } |
| ``` |
| |
