Update README.md
Browse files
README.md
CHANGED
|
@@ -62,9 +62,10 @@ Find more details about the bidirectional architecture and training recipe in ou
|
|
| 62 |
| --------------------- | -------------------------------------- | ----------------------------------- |
|
| 63 |
| **Type** | Late interaction (per-token vectors) | Dense bi-encoder (single vector) |
|
| 64 |
| **Total parameters** | ~353M | ~354M |
|
| 65 |
-
| **
|
|
|
|
| 66 |
| **Context length** | 32,768 tokens | 32,768 tokens |
|
| 67 |
-
| **Vocabulary size** | 64,402 |
|
| 68 |
| **Output** | 128-dim per token | 1024-dim CLS vector |
|
| 69 |
| **Similarity** | MaxSim | Cosine |
|
| 70 |
| **Training precision**| BF16 | BF16 |
|
|
@@ -74,6 +75,8 @@ Find more details about the bidirectional architecture and training recipe in ou
|
|
| 74 |
|
| 75 |
**Supported languages:** English, Spanish, German, French, Italian, Portuguese, Arabic, Swedish, Norwegian, Japanese, Korean.
|
| 76 |
|
|
|
|
|
|
|
| 77 |
```text
|
| 78 |
ColBERT(
|
| 79 |
(0): Transformer({'max_seq_length': 511, 'do_lower_case': False}) with Transformer model: Lfm2BidirectionalModel
|
|
@@ -228,16 +231,16 @@ We highlight (= bold) the best bi-encoder and best late retriever for each langu
|
|
| 228 |
|
| 229 |
| Model | Type | AVG | ar | de | en | es | fr | it | ja | ko | no | pt | sv |
|
| 230 |
| --- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
|
| 231 |
-
|
|
| 232 |
-
|
|
| 233 |
-
|
|
| 234 |
-
|
|
| 235 |
-
|
|
| 236 |
-
|
|
| 237 |
-
|
|
| 238 |
-
|
|
| 239 |
-
|
|
| 240 |
-
|
|
| 241 |
|
| 242 |
### MKQA-11 — Recall@20
|
| 243 |
|
|
@@ -245,16 +248,16 @@ We highlight (= bold) the best bi-encoder and best late retriever for each langu
|
|
| 245 |
|
| 246 |
| Model | Type | AVG | ar | de | en | es | fr | it | ja | ko | no | pt | sv |
|
| 247 |
| --- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
|
| 248 |
-
|
|
| 249 |
-
|
|
| 250 |
-
|
|
| 251 |
-
|
|
| 252 |
-
|
|
| 253 |
-
|
|
| 254 |
-
|
|
| 255 |
-
|
|
| 256 |
-
|
|
| 257 |
-
|
|
| 258 |
|
| 259 |
## 📬 Contact
|
| 260 |
|
|
|
|
| 62 |
| --------------------- | -------------------------------------- | ----------------------------------- |
|
| 63 |
| **Type** | Late interaction (per-token vectors) | Dense bi-encoder (single vector) |
|
| 64 |
| **Total parameters** | ~353M | ~354M |
|
| 65 |
+
| **Backbone** | [LFM2.5-350M-Base](https://huggingface.co/LiquidAI/LFM2.5-350M-Base) + bi-directional patches | [LFM2.5-350M-Base](https://huggingface.co/LiquidAI/LFM2.5-350M-Base) + bi-directional patches |
|
| 66 |
+
| **Layers** | 17 (10 conv + 6 attn + 1 dense) | 17 (10 conv + 6 attn + 1 pool) |
|
| 67 |
| **Context length** | 32,768 tokens | 32,768 tokens |
|
| 68 |
+
| **Vocabulary size** | 64,402 | 65,536 |
|
| 69 |
| **Output** | 128-dim per token | 1024-dim CLS vector |
|
| 70 |
| **Similarity** | MaxSim | Cosine |
|
| 71 |
| **Training precision**| BF16 | BF16 |
|
|
|
|
| 75 |
|
| 76 |
**Supported languages:** English, Spanish, German, French, Italian, Portuguese, Arabic, Swedish, Norwegian, Japanese, Korean.
|
| 77 |
|
| 78 |
+
|
| 79 |
+
|
| 80 |
```text
|
| 81 |
ColBERT(
|
| 82 |
(0): Transformer({'max_seq_length': 511, 'do_lower_case': False}) with Transformer model: Lfm2BidirectionalModel
|
|
|
|
| 231 |
|
| 232 |
| Model | Type | AVG | ar | de | en | es | fr | it | ja | ko | no | pt | sv |
|
| 233 |
| --- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
|
| 234 |
+
| **LiquidAI/LFM2.5-ColBERT-350M** | late | **0.605** | **0.551** | **0.606** | 0.687 | **0.607** | **0.622** | **0.606** | **0.614** | **0.590** | **0.570** | **0.613** | **0.586** |
|
| 235 |
+
| **LiquidAI/LFM2.5-Embedding-350M** | dense | **0.577** | **0.529** | **0.581** | 0.644 | **0.581** | **0.592** | **0.583** | **0.575** | **0.563** | **0.557** | **0.581** | **0.566** |
|
| 236 |
+
| Qwen/Qwen3-Embedding-0.6B | dense | 0.556 | 0.514 | 0.560 | 0.649 | 0.568 | 0.565 | 0.565 | 0.551 | 0.530 | 0.516 | 0.571 | 0.525 |
|
| 237 |
+
| LiquidAI/LFM2-ColBERT-350M | late | 0.540 | 0.491 | 0.563 | 0.661 | 0.563 | 0.564 | 0.543 | 0.557 | 0.527 | 0.449 | 0.547 | 0.480 |
|
| 238 |
+
| Alibaba-NLP/gte-multilingual-base | dense | 0.528 | 0.477 | 0.523 | 0.624 | 0.537 | 0.542 | 0.528 | 0.511 | 0.494 | 0.516 | 0.534 | 0.526 |
|
| 239 |
+
| lightonai/GTE-ModernColBERT-v1 | late | 0.489 | 0.309 | 0.499 | 0.680 | 0.525 | 0.546 | 0.516 | 0.459 | 0.368 | 0.465 | 0.530 | 0.483 |
|
| 240 |
+
| lightonai/LateOn | late | 0.484 | 0.307 | 0.505 | **0.690** | 0.531 | 0.537 | 0.514 | 0.442 | 0.326 | 0.465 | 0.533 | 0.475 |
|
| 241 |
+
| lightonai/DenseOn | dense | 0.432 | 0.178 | 0.474 | **0.676** | 0.496 | 0.520 | 0.487 | 0.378 | 0.197 | 0.422 | 0.493 | 0.433 |
|
| 242 |
+
| Alibaba-NLP/gte-modernbert-base | dense | 0.383 | 0.112 | 0.449 | 0.666 | 0.448 | 0.475 | 0.408 | 0.275 | 0.180 | 0.376 | 0.431 | 0.391 |
|
| 243 |
+
| BAAI/bge-large-en-v1.5 | dense | 0.359 | 0.059 | 0.419 | 0.642 | 0.445 | 0.475 | 0.431 | 0.198 | 0.132 | 0.358 | 0.434 | 0.353 |
|
| 244 |
|
| 245 |
### MKQA-11 — Recall@20
|
| 246 |
|
|
|
|
| 248 |
|
| 249 |
| Model | Type | AVG | ar | de | en | es | fr | it | ja | ko | no | pt | sv |
|
| 250 |
| --- | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
|
| 251 |
+
| **LiquidAI/LFM2.5-ColBERT-350M** | late | **0.694** | **0.608** | **0.709** | 0.748 | **0.711** | **0.715** | **0.707** | **0.703** | **0.640** | **0.689** | **0.703** | **0.700** |
|
| 252 |
+
| **LiquidAI/LFM2.5-Embedding-350M** | dense | **0.691** | **0.610** | **0.709** | 0.738 | **0.708** | **0.715** | **0.703** | **0.685** | **0.630** | 0.691 | **0.710** | **0.708** |
|
| 253 |
+
| Alibaba-NLP/gte-multilingual-base | dense | 0.675 | 0.567 | 0.692 | 0.741 | 0.705 | 0.703 | 0.697 | 0.655 | 0.563 | **0.698** | 0.700 | 0.699 |
|
| 254 |
+
| LiquidAI/LFM2-ColBERT-350M | late | 0.646 | 0.554 | 0.696 | 0.754 | **0.711** | 0.710 | 0.667 | 0.658 | 0.558 | 0.541 | 0.669 | 0.589 |
|
| 255 |
+
| Qwen/Qwen3-Embedding-0.6B | dense | 0.638 | 0.520 | 0.671 | 0.723 | 0.678 | 0.672 | 0.671 | 0.635 | 0.543 | 0.620 | 0.667 | 0.620 |
|
| 256 |
+
| lightonai/GTE-ModernColBERT-v1 | late | 0.459 | 0.092 | 0.532 | 0.754 | 0.552 | 0.615 | 0.510 | 0.275 | 0.166 | 0.503 | 0.524 | 0.524 |
|
| 257 |
+
| lightonai/LateOn | late | 0.454 | 0.157 | 0.492 | **0.755** | 0.537 | 0.577 | 0.481 | 0.316 | 0.209 | 0.472 | 0.502 | 0.501 |
|
| 258 |
+
| lightonai/DenseOn | dense | 0.435 | 0.165 | 0.482 | **0.751** | 0.491 | 0.553 | 0.457 | 0.325 | 0.222 | 0.438 | 0.443 | 0.453 |
|
| 259 |
+
| BAAI/bge-large-en-v1.5 | dense | 0.413 | 0.133 | 0.471 | 0.748 | 0.450 | 0.531 | 0.461 | 0.208 | 0.172 | 0.456 | 0.443 | 0.467 |
|
| 260 |
+
| Alibaba-NLP/gte-modernbert-base | dense | 0.295 | 0.060 | 0.333 | 0.736 | 0.273 | 0.417 | 0.291 | 0.100 | 0.052 | 0.332 | 0.326 | 0.330 |
|
| 261 |
|
| 262 |
## 📬 Contact
|
| 263 |
|