Add MargEmbed-Akan

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

1_Pooling/config.json

@@ -0,0 +1,5 @@
+{
+    "embedding_dimension": 1024,
+    "pooling_mode": "mean",
+    "include_prompt": true
+}

2_Normalize/config.json

@@ -0,0 +1,4 @@
+{
+    "module_input_name": "sentence_embedding",
+    "module_output_name": "sentence_embedding"
+}

README.md

@@ -0,0 +1,37 @@
+---
+library_name: sentence-transformers
+base_model: McGill-NLP/AfriE5-Large-instruct
+language: [tw, en]
+license: mit
+tags: [sentence-transformers, twi, akan, cross-lingual-retrieval, query-encoder, knowledge-distillation]
+---
+# MargEmbed-Akan
+
+Query encoder for **Twi → English retrieval**. **Encode questions with this model and passages with the original
+`McGill-NLP/AfriE5-Large-instruct`** (the passage space is unchanged).
+
+```python
+from sentence_transformers import SentenceTransformer
+q_model = SentenceTransformer("DexelLabs/MargEmbed-Akan")
+p_model = SentenceTransformer("McGill-NLP/AfriE5-Large-instruct")
+prompt = "Instruct: Given a question, retrieve Wikipedia passages that answer the question\nQuery: "
+q = q_model.encode(["Kasa bɛn na nnipa dodow no ara ka wɔ Uganda?"], prompt=prompt, normalize_embeddings=True)
+p = p_model.encode(["Languages of Uganda. In Uganda the most spoken local language is Luganda ..."], normalize_embeddings=True)
+print(q @ p.T)
+```
+
+## Training
+English questions (Natural Questions, TriviaQA, SQuAD, GooAQ; sampled to favour low keyword overlap) translated to
+Asante Twi with `facebook/nllb-200-3.3B` and kept only if their back-translation matches the original. The base model,
+given the **English** question, scores the answer passage and 7 hard negatives; this model learns to
+reproduce those scores from the **Twi** question (listwise KL distillation; passages frozen), plus a light Twi–English
+alignment term on the GhanaNLP corpus. Final weights = 1.0 × trained + 0.0 × base.
+**No AfriQA data was used for training or model selection.** NLLB-200 is CC BY-NC 4.0: research use.
+
+## AfriQA Twi test (nDCG@10, strict; passages encoded by base AfriE5)
+| setting                |   base |   trained |   gemini_ref |
+|:-----------------------|-------:|----------:|-------------:|
+| Twi→Eng                |   67.5 |      71.7 |         68   |
+| Twi (standard ɛ/ɔ)→Eng |   68.3 |      72   |         70   |
+| Twi (no ɛ/ɔ)→Eng       |   63.4 |      70.2 |         63   |
+| Eng→Eng (ref)          |   87.7 |      87.3 |         89.6 |

config.json

@@ -0,0 +1,30 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "XLMRobertaModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "is_decoder": false,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "xlm-roberta",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "output_past": true,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.20.2",
+  "type_vocab_size": 1,
+  "use_cache": false,
+  "vocab_size": 250002
+}

config_sentence_transformers.json

@@ -0,0 +1,14 @@
+{
+  "__version__": {
+    "pytorch": "2.14.1+cu130",
+    "sentence_transformers": "6.2.0",
+    "transformers": "5.20.2"
+  },
+  "default_prompt_name": null,
+  "model_type": "SentenceTransformer",
+  "prompts": {
+    "document": "",
+    "query": ""
+  },
+  "similarity_fn_name": "cosine"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f1f2bb6f1cd4b41dea13cd000009b9cea1ea860d7f36c5f75c016e5b1d89da3
+size 1119826072

modules.json

@@ -0,0 +1,20 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.base.modules.transformer.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.sentence_transformer.modules.pooling.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.base.modules.normalize.Normalize"
+  }
+]

sentence_bert_config.json

@@ -0,0 +1,10 @@
+{
+    "transformer_task": "feature-extraction",
+    "modality_config": {
+        "text": {
+            "method": "forward",
+            "method_output_name": "last_hidden_state"
+        }
+    },
+    "module_output_name": "token_embeddings"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:883b037111086fd4dfebbbc9b7cee11e1517b5e0c0514879478661440f137085
+size 17082987

tokenizer_config.json

@@ -0,0 +1,24 @@
+{
+  "add_prefix_space": true,
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "extra_special_tokens": [],
+  "is_local": true,
+  "local_files_only": false,
+  "mask_token": "<mask>",
+  "max_length": 512,
+  "model_max_length": 512,
+  "pad_to_multiple_of": null,
+  "pad_token": "<pad>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "sep_token": "</s>",
+  "stride": 0,
+  "tokenizer_class": "XLMRobertaTokenizer",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "<unk>"
+}