How model names are cleaned (·Q4 suffix)
Abliteration auto-cleans messy GGUF filenames into readable names, keeping the quantization as a suffix.
Short answer
Abliteration auto-cleans messy GGUF filenames into readable names, keeping the quantization as a suffix. Llama-3.2-3B-Instruct-Q4_K_M.gguf becomes Llama 3.2 3B Instruct · Q4.
Detailed explanation
GGUF filenames from Hugging Face tend to look like:
Llama-3.2-3B-Instruct-Q4_K_M.ggufqwen2.5-7b-instruct-q4_k_m.ggufMeta-Llama-3.1-8B-Instruct-abliterated.Q4_K_M.gguf
These are functional but ugly in a list. Abliteration’s auto-clean does five things:
- Removes
.ggufextension - Extracts the quantization suffix (Q4_K_M, Q8_0, IQ4_XS, F16, etc.)
- Replaces
-and_with spaces - Collapses multiple spaces
- Appends the quant back as
· Q4(or· Q8,· IQ4,· F16, etc.)
Result: Llama 3.2 3B Instruct · Q4. Same information, human-readable.
Preserving quantization matters because two versions of the same model at different quants would otherwise get the same cleaned name and conflict. Llama 3.2 3B Instruct · Q4 and Llama 3.2 3B Instruct · Q8 are clearly different in the list.
Short quant mapping:
| Full quant | Displayed as |
|---|---|
| Q4_K_M, Q4_K_S, Q4_0 | Q4 |
| Q5_K_M, Q5_K_S | Q5 |
| Q6_K | Q6 |
| Q8_0 | Q8 |
| Q3_K_M, Q3_K_L | Q3 |
| Q2_K | Q2 |
| IQ4_XS, IQ4_NL | IQ4 |
| IQ3_M, IQ3_XS | IQ3 |
| IQ2_M | IQ2 |
| F16, BF16 | F16 |
The quant is read from the GGUF metadata (general.file_type) when possible - this is more reliable than parsing the filename. Filename parsing is the fallback.
Why it matters
Clean names make the Models list scannable at a glance, and the quant suffix means you always know which version you’re looking at without opening file details.
In Abliteration specifically
Auto-clean happens on model add. The cleaned name is stored in manifest.json (in the models folder) and shown everywhere in the UI - Models list, chat model selector, fit-check dialog, drawer titles.
The Edit dialog (⋯ → Edit) lets you switch between three name modes:
- Auto-cleaned - the default (e.g.
Llama 3.2 3B Instruct · Q4) - Original filename - the raw GGUF filename (
Llama-3.2-3B-Instruct-Q4_K_M) - Custom - anything you type
Cyrillic filenames and other non-Latin scripts are preserved as-is through cleaning (модель_финал.gguf → модель финал · Q4).
Related
- Quantization - what the suffix means
- Edit model dialog - where you switch name modes