---
license: mit
base_model: microsoft/codebert-base
base_model_relation: quantized
library_name: gguf
pipeline_tag: feature-extraction
tags:
  - gguf
  - embeddings
  - code
  - llama.cpp
---

# CodeBERT base — GGUF Q8_0

[`microsoft/codebert-base`](https://huggingface.co/microsoft/codebert-base) converted to GGUF at Q8_0 for
llama.cpp's `llama-server --embeddings`.

| | |
|---|---|
| File | `codebert-base-Q8_0.gguf` (135 MB) |
| SHA-256 | `01a85fb726bf4dc063aaac34b33fde08f991b96b3f9f76ff10e28fa2269df5ea` |
| Source revision | `99d7ef814601faaf7bdc2f774ffa7dade4f4d828` (safetensors) |
| Context | 512 tokens |
| Pooling | mean (stored in the file, no `--pooling` flag needed) |
| Dimensions | 768 |

```sh
llama-server -m codebert-base-Q8_0.gguf --embeddings --port 8080
curl -s localhost:8080/v1/embeddings -H 'Content-Type: application/json' -d '{"input":["def f(): pass"]}'
```

## How it was converted

`convert_hf_to_gguf.py` (llama.cpp, 2026-10-01) with `--outtype q8_0`, after two additions to the source
directory:

- a `tokenizer.json`, written by `RobertaTokenizerFast.save_pretrained`. The upstream repo ships only
  `vocab.json` + `merges.txt`, and without `tokenizer.json` the converter falls back to a WordPiece
  vocabulary and tokenizes code wrongly;
- a sentence-transformers `modules.json` + `1_Pooling/config.json` declaring mean pooling, so the pooling
  type is written into the GGUF.

## Verification

40 random code/console/output blocks (first 600 characters), embedded by this file on llama-server (CPU) and
by the original model in PyTorch fp32 with masked mean pooling over the last hidden state. Token ids are
identical; cosine similarity min 0.9998, median 0.99994.

The weights and their licence (MIT) are Microsoft's; see the [CodeBERT paper](https://arxiv.org/abs/2002.08155).