Llama-3.1-8B-ERP-WHW-GGUF

Llama-3.1-8B-ERP-WHW-GGUF

Built with Llama.

A prompt-configured build of Meta-Llama-3.1-8B-Instruct (Q4_K_M GGUF) that answers internal ERP process questions in a fixed WHW — What / How / Why format and refuses everything else.

The weights are unmodified. This repository contains no fine-tuning, no LoRA, and no continued pretraining. What is packaged here is the base Q4_K_M quantization plus a system prompt, chat template, and decoding parameters. Anything the model "knows" about ERP comes from Llama 3.1's pretraining, not from this repo.

Files

Usage — Ollama

# Fetch the repo (needs git-lfs for the .gguf)
git lfs install
git clone https://huggingface.co/hanzag/Llama-3.1-8B-ERP-WHW-GGUF
cd Llama-3.1-8B-ERP-WHW-GGUF

ollama create erp-whw -f Modelfile
ollama run erp-whw "How do I raise a purchase requisition?"

Or straight from the Hub, without cloning:

ollama run hf.co/hanzag/Llama-3.1-8B-ERP-WHW-GGUF

Note that the direct hf.co/ form loads the GGUF but not the Modelfile — you get the base model's default template and no system prompt. Use ollama create if you want the WHW behaviour.

Usage — llama.cpp

llama-cli -m Meta-Llama-3.1-8B-Instruct.Q4_K_M.gguf \
  --system-prompt-file system_prompt.txt \
  --temp 0 -c 4096

Output format

In-scope queries are answered as:

WHAT: <the thing>
HOW:  <the steps>
WHY:  <the rationale>

Out-of-scope queries return a fixed refusal:

WHAT: Out of scope request.
HOW: ACTION BLOCKED: NON-ERP_CONTEXT.
WHY: <one sentence reason>

Parameters

temperature 0 · num_ctx 4096 · stop tokens <|start_header_id|>, <|end_header_id|>, <|eot_id|>

Temperature is pinned at 0 for repeatable answers. It reduces variance; it does not prevent the model from being wrong, and a confidently-worded WHW answer can still be fabricated.

Limitations

• Prompt-level scoping only. The refusal behaviour is instruction-following, not a guardrail. It can be talked around, and it is not a security control.
• No ERP-specific knowledge was added. Answers about a particular ERP deployment, its customizations, or its data will not be reliable. Verify against your own documentation.
• Q4_K_M quantization trades some quality for size versus the full-precision model.
• English only, as configured.

License

Governed by the Llama 3.1 Community License and the Llama 3.1 Acceptable Use Policy. Derivative works must carry the "Built with Llama" notice and keep the Llama name prefix.