---
license: other
language:
- th
- en
tags:
- gguf
- tinymind
- qlora
- evidence-gated
- local-llm
pipeline_tag: text-generation
library_name: gguf
base_model: mistralai/Mistral-Nemo-Instruct-2407
---

# TinyMind GGUF-X PureCode

This repository contains the TinyMind GGUF-X PureCode runtime package.

GGUF-X is a TinyMind packaging format: a GGUF v3-compatible binary runtime plus sidecar evidence, decoding policy, purity lineage, and claim gates. It is designed to keep the runtime compatible with GGUF loaders while attaching the evidence needed to avoid fake claims.

## What Is Included

- `Modelfile.evo`: Ollama-compatible runtime profile.
- `tinymind_ggufx_spec.json`: TinyMind GGUF-X sidecar spec.
- `gguf_evo_upgrade_manifest.json`: hashes, lineage, runtime settings, and promotion gates.
- `gguf_evo_eval_prompts.jsonl`: smoke eval prompts.
- `README_GGUF_X.md`: concise format boundary.

## Current Evidence

- Source GGUF size: `15.7098 GB`
- Source GGUF SHA-256: `b8707e57f676d8dd1b80f623b45200cc92e6966b0e95275e606f412095a49fde`
- Adapter training evidence present: `True`
- Adapter eval loss: `2.808002471923828`
- Adapter perplexity: `16.576772556518506`
- Purity density: `0.7211701159881736`
- Dominant domain share: `0.17511939959062997`
- Runtime context requested: `65536`

## Claim Boundary

- Custom TinyMind format pack: `True`
- GGUF v3 compatibility: `True`
- Better-than-GGUF-v3 claim: `False`
- Weights-better-than-source claim: `False`

The current package does not claim that GGUF tensors were merged/retrained inside the binary. Weight-level improvement requires a real adapter-to-GGUF merge/export log and same-prompt baseline comparison.

## Reproduce Locally

```powershell
ollama create tinymind-ggufx-purecode -f .\Modelfile.evo
ollama run tinymind-ggufx-purecode "อธิบายความต่างของ QLoRA กับ GGUF"
```

## Repository

HF repo target: `bbkdevops/tinymind-ggufx-purecode`