From 2fe54c71559e162995489f62a812ba82df1afc4e Mon Sep 17 00:00:00 2001 From: Mattijs Ugen <144798+akaIDIOT@users.noreply.github.com> Date: Fri, 7 Aug 2026 15:37:17 +0200 Subject: [PATCH 1/2] Use new model identifier in embedder test --- asmtransformers/tests/test_embedder.py | 8 ++++---- citatio/README.md | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/asmtransformers/tests/test_embedder.py b/asmtransformers/tests/test_embedder.py index 2a19ad7..c946536 100644 --- a/asmtransformers/tests/test_embedder.py +++ b/asmtransformers/tests/test_embedder.py @@ -170,10 +170,10 @@ def test_mean_pool_ignores_padding(): @pytest.mark.skipif(os.environ.get('CI') == 'true', reason="don't run this test on CI") def test_hf_embedding_model_loads_natively(anchor): - embedder = ASMEmbedder.from_pretrained('NetherlandsForensicInstitute/ARM64BERT-embedding') + embedder = ASMEmbedder.from_pretrained('NetherlandsForensicInstitute/Multilingual-ASMBERT') embedding = embedder.encode(anchor, architecture='arm64') - assert np.isclose(embedding.sum(), -0.09272218) - assert np.isclose(embedding.min(), -0.10641833) - assert np.isclose(embedding.max(), 0.116405316) + assert np.isclose(embedding.sum(), -0.0179356) + assert np.isclose(embedding.min(), -0.1113021) + assert np.isclose(embedding.max(), 0.12490158) diff --git a/citatio/README.md b/citatio/README.md index 7ad9da1..477bac0 100644 --- a/citatio/README.md +++ b/citatio/README.md @@ -2,7 +2,7 @@ Citatio 📜 ========== This package provides a REST API to the [`asmtransformers`](../asmtransformers) and its -[`ARM64BERT-embedding`](https://huggingface.co/NetherlandsForensicInstitute/ARM64BERT-embedding/) model, +[`ASMBERT-embedding`](https://huggingface.co/NetherlandsForensicInstitute/ARM64BERT-embedding/) model, bridging Ghidra to a search index to find similar functions in ARM64 binaries. See [`sententia`](../sententia) for the Ghidra plugin that uses this REST API to both add functions to the search index and search for possible function names / labels based on vector similarity. From 17d46687868ce7e7a9386737038cd033a09b5d65 Mon Sep 17 00:00:00 2001 From: Mattijs Ugen <144798+akaIDIOT@users.noreply.github.com> Date: Fri, 7 Aug 2026 15:40:03 +0200 Subject: [PATCH 2/2] Replace all references to the old model identifier that are not explicitly about the old model --- asmtransformers/README.md | 8 ++++---- asmtransformers/asmtransformers/models/asmsentencebert.py | 2 +- asmtransformers/asmtransformers/models/embedder.py | 2 +- citatio/README.md | 6 +++--- 4 files changed, 9 insertions(+), 9 deletions(-) diff --git a/asmtransformers/README.md b/asmtransformers/README.md index c30ed28..a932276 100644 --- a/asmtransformers/README.md +++ b/asmtransformers/README.md @@ -126,8 +126,8 @@ by Wang et al. in the jTrans paper (referred to above), is implicitly included i JUMP-token, the correct token to predict is the correct jump address. This is the token index of the place the code was supposed to jump to, materialised in the vocabulary as JUMP_ADDR_n `(n = 1, len(max_token_lenght))`. -You can also find our monolingual, pretrained only ARM64BERT model on Huggingface: -NetherlandsForensicInstitute/ARM64Bert +You can also find our monolingual, pretrained only ARM64BERT model on Hugging Face: +[NetherlandsForensicInstitute/ARM64Bert](https://huggingface.co/NetherlandsForensicInstitute/ARM64Bert) Finetuning ---------- @@ -152,8 +152,8 @@ to train the model. Our multilingual model can be found on [Hugging Face](https://huggingface.co/NetherlandsForensicInstitute/Multilingual-ASMBERT) -You can also find our monolingual, ARM64BERT semantic search model on Hugging Face: -NetherlandsForensicInstitute/ARM64bert-embedding +You can also find our monolingual, ARM64BERT-embedding semantic search model on Hugging Face: +[NetherlandsForensicInstitute/ARM64bert-embedding](https://huggingface.co/NetherlandsForensicInstitute/ARM64bert-embedding) Evaluation ---------- diff --git a/asmtransformers/asmtransformers/models/asmsentencebert.py b/asmtransformers/asmtransformers/models/asmsentencebert.py index d7f7ce0..06af6f5 100644 --- a/asmtransformers/asmtransformers/models/asmsentencebert.py +++ b/asmtransformers/asmtransformers/models/asmsentencebert.py @@ -10,7 +10,7 @@ class ASMTransformerModule(Module): - """Minimal sentence-transformers module for ARM64BERT finetuning.""" + """Minimal sentence-transformers module for ASMBERT finetuning.""" def __init__( self, diff --git a/asmtransformers/asmtransformers/models/embedder.py b/asmtransformers/asmtransformers/models/embedder.py index ec90b59..de36e8a 100644 --- a/asmtransformers/asmtransformers/models/embedder.py +++ b/asmtransformers/asmtransformers/models/embedder.py @@ -8,7 +8,7 @@ class ASMEmbedder: - """Native inference wrapper for ARM64BERT-style embedding checkpoints.""" + """Native inference wrapper for ASMBERT-style embedding checkpoints.""" def __init__(self, model, tokenizer, *, device=None, normalize_embeddings=True): self.model = model diff --git a/citatio/README.md b/citatio/README.md index 477bac0..867e028 100644 --- a/citatio/README.md +++ b/citatio/README.md @@ -2,7 +2,7 @@ Citatio 📜 ========== This package provides a REST API to the [`asmtransformers`](../asmtransformers) and its -[`ASMBERT-embedding`](https://huggingface.co/NetherlandsForensicInstitute/ARM64BERT-embedding/) model, +[`Multilingual-ASMBERT`](https://huggingface.co/NetherlandsForensicInstitute/Multilingual-ASMBERT/) model, bridging Ghidra to a search index to find similar functions in ARM64 binaries. See [`sententia`](../sententia) for the Ghidra plugin that uses this REST API to both add functions to the search index and search for possible function names / labels based on vector similarity. @@ -12,13 +12,13 @@ Configuration and runtime The citatio REST API takes 3 configuration options: -- The model to be used for embedding (currently, only `NetherlandsForensicInstitute/ARM64BERT-embedding` is supported); +- The model to be used for embedding (currently, only `NetherlandsForensicInstitute/Multilingual-ASMBERT` is supported); - The authentication modes to support, any of `anonymous`, `client_supplied` and `oidc`; - The database to store both assembly and embeddings in, either SQLite+sqlitevec or PostgreSQL+pgvector; Citatio uses [confidence](https://github.com/NetherlandsForensicInstitute/confidence/) to read configuration, so both strategically placed files and environment variables are supported: -- `CITATIO_MODEL_HF` or `CITATIO_MODEL_PATH`: a local path or huggingface model name (though again, currently only the `ARM64BERT-embedding` model is supported); +- `CITATIO_MODEL_HF` or `CITATIO_MODEL_PATH`: a local path or huggingface model name (though again, currently only the `NetherlandsForensicInstitute/Multilingual-ASMBERT` model is supported); - `CITATIO_AUTH_ANONYMOUS` (allowing operation without identifying a user) and `CITATIO_AUTH_CLIENT__SUPPLIED` (enabling a client to supply a user identity in a request body) can be set to `true` to enable them, OIDC configuration requires at least four values, see below.