Model licenses¶
Lasagna ships no model weights. Every embedding, reranking and sparse-encoding model is downloaded from its upstream host (usually the Hugging Face Hub) the first time you use it, and each carries its own license, independent of Lasagna's Apache 2.0.
Most are permissive. Four are not, and they are selectable on purpose: Lasagna exists to compare RAG components, and excluding well-known models would undercut that. What matters is that you know which one you picked.
Short version: for research, evaluation and benchmarking, everything below is fine. Before putting a configuration into a commercial product, check this page: four models forbid that use, or do not say they allow it.
Restricted: research and evaluation only¶
| Model | License | Where it appears | Notes |
|---|---|---|---|
naver/splade-cocondenser-ensembledistil |
CC-BY-NC-SA-4.0 | Default SPLADE sparse encoder | Non-commercial, share-alike. Sparse retrieval is opt-in, so this loads only when you enable it. |
jinaai/jina-reranker-v2-base-multilingual |
CC-BY-NC-4.0 | Reranker option | Non-commercial. Jina AI sells a commercial license separately. |
BAAI/bge-reranker-v2-gemma |
Gemma Terms of Use | Reranker option | Built on Gemma; inherits use restrictions. Not an OSI-approved open source license. |
castorini/monot5-base-msmarco-10k |
none stated | Default monot5 reranker |
The model card states no license. T5-base (Apache-2.0) fine-tuned on MS MARCO, whose terms allow non-commercial research use; ask the authors before using it in a product. Loads only when a sweep or setting names monot5. |
Selecting any of these logs a WARNING the first time the model loads, naming
the license and linking the model card. The warning is advisory. Lasagna will
not stop you, because for research use there is nothing to stop.
Permissive: commercial use allowed¶
| Model | License |
|---|---|
BAAI/bge-m3 |
MIT |
BAAI/bge-large-en-v1.5 |
MIT |
BAAI/bge-base-en-v1.5 |
MIT |
BAAI/bge-small-en-v1.5 |
MIT |
BAAI/bge-reranker-base |
MIT |
BAAI/bge-reranker-large |
MIT |
BAAI/bge-reranker-v2-m3 |
Apache-2.0 |
intfloat/e5-small-v2 |
MIT |
intfloat/e5-base-v2 |
MIT |
intfloat/e5-large-v2 |
MIT |
intfloat/multilingual-e5-large |
MIT |
cross-encoder/ms-marco-MiniLM-L-6-v2 |
Apache-2.0 |
cross-encoder/ms-marco-MiniLM-L-12-v2 |
Apache-2.0 |
cross-encoder/ms-marco-TinyBERT-L-2-v2 |
Apache-2.0 |
cross-encoder/ms-marco-electra-base |
Apache-2.0 |
mixedbread-ai/mxbai-rerank-large-v1 |
Apache-2.0 |
sentence-transformers/all-MiniLM-L6-v2 |
Apache-2.0 |
colbert-ir/colbertv2.0 |
MIT |
ms-marco-TinyBERT-L-2-v2 (FlashRank's default, from prithivida/flashrank) |
CC-BY-SA-4.0: attribution, and share-alike for adaptations |
A model named without its organization (all-MiniLM-L6-v2, the speed
profile's embedder) is sentence-transformers' own, and is looked up as such.
tests/test_model_licenses.py fails when the default model of a built-in
component, or a profile's model, has no entry here.
Checking programmatically¶
from lasagna.core.model_licenses import describe, is_commercial_use_restricted, restricted_models
describe("naver/splade-cocondenser-ensembledistil")
# LicenseInfo(spdx='CC-BY-NC-SA-4.0', commercial_use=False, url='https://huggingface.co/...', note=...)
is_commercial_use_restricted("BAAI/bge-m3") # False
sorted(restricted_models()) # every non-commercial model
Useful in a pre-deploy check:
from lasagna.core.config import get_rag_config
from lasagna.core.model_licenses import is_commercial_use_restricted
config = get_rag_config()
for model in (config.embedding_model, config.reranker_model):
if is_commercial_use_restricted(model):
raise SystemExit(f"{model} is not licensed for commercial use")
Bringing your own model¶
Lasagna does not restrict which model identifier you configure. Models absent
from the table load normally and log a DEBUG note that their license was not
checked. Absence means uncataloged, not unrestricted.
If you add a model to the configurator, add it to
src/lasagna/core/model_licenses.py
too. tests/test_model_licenses.py fails the build when the UI offers a model
the table does not cover, which is what keeps this page honest.
Caveats¶
This table is a convenience, not legal advice, and licenses change. The model card linked from each entry is authoritative; confirm there before relying on anything here.
Lasagna also redistributes no benchmark corpora; evaluation datasets are fetched from their official sources and carry their own terms.