README.md

---
base_model: ornith-ai/Ornith-1.5-9B
datasets:
- guell00/fds
library_name: gguf
pipeline_tag: text-generation
license: mit
tags:
- gguf
- llama.cpp
- code
- coding
- imatrix
- qwen3.5
- ornith
- lora
---

# VELUM-Coder

**VELUM-Coder** is a coding-focused fine-tune derived from
[`ornith-ai/Ornith-1.5-9B`](https://huggingface.co/ornith-ai/Ornith-1.5-9B).

This repository contains GGUF quantizations generated from the completed recovery adapter.

- Source adapter: `ornith_fds_recovered_ckpt250`
- Dataset: `guell00/fds`
- llama.cpp commit: `f280b26983ad0fdb705a0d9ebf0503e76f2899b0`
- Generated: 2026-08-25 01:57 UTC

## Quantization policy

- `Q8_0` is intentionally generated **without** an importance matrix.
- Every Q4, Q3, Q2 and IQ1 build is invoked with the same domain-specific **imatrix**.
- No Q5 or Q6 files are produced.
- The extreme `IQ1_M` build keeps the token embedding and output tensors at `Q8_0` while the body is quantized as `IQ1_M`.
- The imatrix is generated from code/conversation samples from `guell00/fds`.
- `<think>...</think>` blocks are removed from assistant calibration text so calibration emphasizes direct answers and code.

## Recommended files

**Best default:** `VELUM-Coder-Q4_K_M.gguf`

**Higher quality:** `VELUM-Coder-Q8_0.gguf`

**Lower memory:** try `VELUM-Coder-IQ3_M.gguf` or `VELUM-Coder-IQ2_M.gguf`.

**Extreme experiment:** `VELUM-Coder-IQ1_M.gguf`. This is not expected to preserve Q4-level quality. Embedding/output protection exists only to reduce catastrophic collapse.

## Files

| Status | File | Quant | Imatrix | Protection | Notes |
|---|---|---|---|---|---|
| ✅ | `VELUM-Coder-Q8_0.gguf` | `Q8_0` | No | Normal | Highest-quality quantized build; intentionally no imatrix. |
| ✅ | `VELUM-Coder-Q4_K_M.gguf` | `Q4_K_M` | Yes | Normal | Recommended default Q4. |
| ✅ | `VELUM-Coder-Q4_K_S.gguf` | `Q4_K_S` | Yes | Normal | Smaller Q4 K-quant. |
| ✅ | `VELUM-Coder-IQ4_XS.gguf` | `IQ4_XS` | Yes | Normal | Compact 4-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ4_NL.gguf` | `IQ4_NL` | Yes | Normal | Non-linear 4-bit I-Quant. |
| ✅ | `VELUM-Coder-Q4_1.gguf` | `Q4_1` | Yes | Normal | Legacy Q4 variant. |
| ✅ | `VELUM-Coder-Q4_0.gguf` | `Q4_0` | Yes | Normal | Legacy compact Q4. |
| ✅ | `VELUM-Coder-Q3_K_L.gguf` | `Q3_K_L` | Yes | Normal | Largest Q3 K-quant. |
| ✅ | `VELUM-Coder-Q3_K_M.gguf` | `Q3_K_M` | Yes | Normal | Balanced Q3 K-quant. |
| ✅ | `VELUM-Coder-Q3_K_S.gguf` | `Q3_K_S` | Yes | Normal | Smaller Q3 K-quant. |
| ✅ | `VELUM-Coder-IQ3_M.gguf` | `IQ3_M` | Yes | Normal | Medium 3-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ3_S.gguf` | `IQ3_S` | Yes | Normal | Small 3-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ3_XS.gguf` | `IQ3_XS` | Yes | Normal | Extra-small 3-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ3_XXS.gguf` | `IQ3_XXS` | Yes | Normal | Extremely compact 3-bit I-Quant. |
| ✅ | `VELUM-Coder-Q2_K.gguf` | `Q2_K` | Yes | Normal | Aggressive Q2 K-quant. |
| ✅ | `VELUM-Coder-Q2_K_S.gguf` | `Q2_K_S` | Yes | Normal | Small Q2 K-quant. |
| ✅ | `VELUM-Coder-IQ2_M.gguf` | `IQ2_M` | Yes | Normal | Higher-quality 2-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ2_S.gguf` | `IQ2_S` | Yes | Normal | 2-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ2_XS.gguf` | `IQ2_XS` | Yes | Normal | Extra-small 2-bit I-Quant. |
| ✅ | `VELUM-Coder-IQ2_XXS.gguf` | `IQ2_XXS` | Yes | Normal | Extremely compact 2-bit I-Quant. |
| ⏳ | `VELUM-Coder-Q2_0.gguf` | `Q2_0` | Yes | Normal | 2.25 bpw Q2_0. |
| ⏳ | `VELUM-Coder-TQ2_0.gguf` | `TQ2_0` | Yes | Normal | Experimental ternary ~2-bit build. |
| ⏳ | `VELUM-Coder-IQ1_M.gguf` | `IQ1_M` | Yes | Q8_0 embed/output | Extreme build. IQ1_M body + Q8_0 embedding/output protection. |

## Importance matrix

The generated importance matrix is uploaded as:

`imatrix/VELUM-Coder-imatrix.gguf`

The calibration text itself is not uploaded; it is only an intermediate generated from `guell00/fds`.

## llama.cpp

Example:

~~~bash
llama-cli -hf guell00/VELUM-Coder:Q4_K_M -c 16384
~~~

Server:

~~~bash
llama-server -hf guell00/VELUM-Coder:Q4_K_M -c 16384
~~~

Manual download:

~~~bash
hf download guell00/VELUM-Coder VELUM-Coder-Q4_K_M.gguf --local-dir .
~~~

Then:

~~~bash
llama-cli -m VELUM-Coder-Q4_K_M.gguf -c 16384
~~~

## Adapter

The final recovery LoRA adapter is also uploaded under `adapter/` when adapter upload is enabled.

## Notes

Low-bit quantization is destructive. Q2 and especially IQ1_M are experimental options for constrained hardware, not substitutes for Q4/Q8 quality.

GGUFs are produced directly by `llama-quantize` and uploaded after successful quantization. No conversational inference is run by this pipeline.

## Credits

- Ornith authors
- llama.cpp / ggml contributors
- Hugging Face
- Unsloth
- `guell00/fds`