README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen3-0.6B-Base
library_name: llama.cpp
pipeline_tag: feature-extraction
model_type: embedding
quantization:
  - Q4_K_M
tags:
  - embeddings
  - qwen
  - gguf
  - tokenizer
language:
  - en
inference: false

Qwen3-Embedding-0.6B (GGUF) Models

This directory contains GGUF builds of the Qwen3 0.6B embedding model, produced from the upstream base repository Qwen/Qwen3-0.6B-Base (original Hugging Face layout in ../Qwen3-Embedding-0.6B/).

Contents

File	Purpose
qwen3-embedding-0.6b.Q4_K_M.gguf	Quantized (Q4_K_M) GGUF for efficient inference.
qwen3-embedding-0.6b-fix.gguf	Same model with explicit sep_token / EOS metadata fix applied.

Special Token Configuration

Extracted from tokenizer_config.json:

"sep_token": "<|endoftext|>",
"sep_token_id": 151643

The model uses <|endoftext|> as both the padding (pad_token) and separator (sep_token). For embedding generation each input text MUST terminate with the separator token (or the converter must auto-append it) to avoid a runtime warning:

[WARNING] At least one last token in strings embedded is not SEP. 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header

Why the Warning Appears

If the GGUF metadata key tokenizer.ggml.add_eos_token is absent or false, llama.cpp will not auto-append the final SEP/EOS token for embedding inputs. Any input string that does not already end with <|endoftext|> triggers the warning and may yield sub‑optimal embeddings (slightly different token boundary semantics).

Fix Implemented

The file qwen3-embedding-0.6b-fix.gguf was regenerated ensuring:

• tokenizer.ggml.add_eos_token = true
• sep_token (<|endoftext|>) retained with id 151643

This makes llama.cpp automatically append the SEP/EOS token when missing, silencing the warning and standardizing embeddings.

Rebuilding From Upstream (Recommended Process)

1. Obtain upstream model:
   • Clone or download Qwen/Qwen3-0.6B-Base (embedding variant directory).
2. Convert to GGUF using the current llama.cpp conversion script:
   • Use the repo's convert_hf_to_gguf.py (it already sets EOS for Qwen tokenizers). Example:

python3 llama.cpp/convert_hf_to_gguf.py \
  --model Qwen3-Embedding-0.6B \
  --outfile qwen3-embedding-0.6b-fix.gguf \
  --ftype q4_k_m

If you previously produced a GGUF that shows the warning, just re-run conversion with an up-to-date llama.cpp checkout. The script internally writes tokenizer.ggml.add_eos_token = true for this tokenizer family.

Post-Conversion Validation

Run a quick embedding call and confirm no warning appears:

./llama.cpp/build/bin/embedding \
  -m models/qwen3-embedding-0.6b-fix.gguf \
  -p "Hello world"

If you still see the warning:

• Confirm the binary was rebuilt after updating sources (make or cmake --build).
• Inspect metadata using a small Python snippet:

from gguf import GGUFReader
r = GGUFReader("models/qwen3-embedding-0.6b-fix.gguf")
for f in r.fields:
    if f.name == "tokenizer.ggml.add_eos_token":
        print("ADD_EOS_TOKEN=", f.parts[-1])

Expected output: ADD_EOS_TOKEN= True

Manual Patch (Fallback Method)

If re-conversion is inconvenient, you can clone metadata and force the flag:

from gguf import GGUFReader, GGUFWriter, constants as C
src = GGUFReader("qwen3-embedding-0.6b.Q4_K_M.gguf")
dst = GGUFWriter("qwen3-embedding-0.6b-fix.gguf", src.architecture)

# Copy all existing fields except override ADD_EOS
for field in src.fields:
    if field.name == C.Keys.Tokenizer.ADD_EOS:
        continue
    dst.add_field(field.name, field.field_type, field.parts)

dst.add_add_eos_token(True)  # set flag

# Copy tensors
for tensor in src.tensors:
    data = tensor.data()
    dst.add_tensor(tensor.name, data, tensor.shape, tensor.tensor_type)

dst.write_header_to_file()
dst.write_kv_data_to_file()
dst.write_tensors_to_file()
dst.close()

After patching, re-run the validation step.

Usage Notes for Embeddings

• Always feed raw text; no special wrapping needed. Auto-SEP happens with the fixed file.
• For batch embeddings, ensure each string ends cleanly (avoid trailing spaces if you rely on identical hashes downstream).
• The dimensionality matches upstream Qwen3-Embedding-0.6B (refer to upstream docs for exact embedding size).

License & Attribution

The original model weights and tokenizer come from the Qwen project (Qwen/Qwen3-0.6B-Base). Review their license and usage terms before redistribution. This README documents conversion adjustments only (metadata EOS flag addition).

Changelog

• Initial addition: added fixed GGUF with tokenizer.ggml.add_eos_token = true to suppress SEP warning.

---

For further improvements (FP16 build, alternative quantization tiers, or batching examples), open an issue or PR in this repo.