---
language:
- en
- de
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
- terraform
- hcl
- iac
- savant
- gguf
- llama-cpp
- ollama
- eschaton
- eu-ai-act
- annex-iv
- narrow-domain
pipeline_tag: text-generation
inference: false
---

# Eschaton Terraform Savant v0.4.19 — GGUF

GGUF-quantized build of [ljawadi/eschaton-terraform-savant-v0.4.19](https://huggingface.co/ljawadi/eschaton-terraform-savant-v0.4.19)
with the LoRA *merged into the base*, so the model loads directly in
`llama.cpp`, Ollama, LM Studio, `text-generation-webui`, and any other
runtime that consumes GGUF — **no GPU required.**

For training methodology, evals, EU AI Act Annex IV compliance pack,
limitations, and citation, see the [main repo](https://huggingface.co/ljawadi/eschaton-terraform-savant-v0.4.19).

## Available quantizations

| File | Size | Use case |
|------|-----:|----------|
| `eschaton-terraform-savant-v0.4.19-Q4_K_M.gguf` | ~4.4 GB | Laptop CPU, 8 GB GPU (RTX 3070 / A2000), Apple Silicon M1/M2 |
| `eschaton-terraform-savant-v0.4.19-Q5_K_M.gguf` | ~5.1 GB | Better quality, 12 GB GPU (RTX 3060 / 4070) |
| `eschaton-terraform-savant-v0.4.19-Q8_0.gguf`   | ~7.5 GB | Near-lossless, 16+ GB GPU / Apple Silicon 16 GB+ |

## Ollama

```bash
ollama run hf.co/ljawadi/eschaton-terraform-savant-v0.4.19-gguf:Q4_K_M
```

Then prompt:

```
>>> Write the complete Terraform HCL code for an aws_s3_bucket 'logs-prod' in eu-central-1
   with versioning, SSE-S3, public-access blocked. Output ONLY HCL in a fenced code block.
```

The model is trained to refuse non-Terraform tasks. By design (see main repo).

## llama.cpp

```bash
./llama-cli -m eschaton-terraform-savant-v0.4.19-Q4_K_M.gguf \
  -p "<|im_start|>user\nWrite the Terraform HCL for an aws_s3_bucket 'logs-prod' \
in eu-central-1 with versioning, SSE-S3, public-access blocked. Output ONLY HCL in \
a fenced code block.<|im_end|>\n<|im_start|>assistant\n" \
  -n 1024 --temp 0 --no-display-prompt
```

## LM Studio / text-generation-webui

Drop any of the `.gguf` files into the models folder and select it from the UI.
Use the **Qwen2.5** chat template (built into both UIs).

## Out of scope — refuses by design

Like the main repo, this savant is **only** for Terraform HCL. Asking it
anything else (python, kubernetes, dockerfile, general questions, even
"explain Terraform CI/CD in GitHub Actions") will trigger the canonical
refusal:

> *"I am a Terraform savant. This task is out of scope. Please ask
> Terraform/HCL questions only."*

That refusal score of **1.00** on the held-out OOS bench + **1.00** on a 10-prompt
adversarial functional test is the auditable evidence for the Annex IV §3
intended-use declaration.

## Contact

[eschatonai.eu](https://eschatonai.eu) *(website may not be reachable yet —
DM [LinkedIn](https://www.linkedin.com/in/lukas-sami-jawadi/))*

---

# 🇩🇪 Deutsche Kurzversion

GGUF-quantisierter Build von
[ljawadi/eschaton-terraform-savant-v0.4.19](https://huggingface.co/ljawadi/eschaton-terraform-savant-v0.4.19).
Die LoRA ist *in die Base gemerged*, sodass das Modell direkt in `llama.cpp`,
Ollama, LM Studio etc. lädt — **kein GPU nötig.**

Für Methodik, Evals, Annex IV Pack und Limitations siehe den
[Haupt-Repo](https://huggingface.co/ljawadi/eschaton-terraform-savant-v0.4.19).

**Drei Quantization-Stufen:** Q4_K_M (Laptop, ~4.4 GB), Q5_K_M (besser, ~5.1 GB),
Q8_0 (near-lossless, ~7.5 GB).

**Ollama:** `ollama run hf.co/ljawadi/eschaton-terraform-savant-v0.4.19-gguf:Q4_K_M`

**Out of scope:** Das Modell lehnt alle Nicht-Terraform-Anfragen ab. Refusal-Score
1.00 perfekt — das ist die auditierbare Evidenz für die Annex-IV-§3-Intended-Use-Deklaration.

**Kontakt:** [eschatonai.eu](https://eschatonai.eu) *(ggf. noch nicht erreichbar —
DM auf [LinkedIn](https://www.linkedin.com/in/lukas-sami-jawadi/))*