Minimize card: remove method and third-party component details (trade secret)
Browse files
README.md
CHANGED
|
@@ -2,36 +2,17 @@
|
|
| 2 |
license: apache-2.0
|
| 3 |
tags:
|
| 4 |
- retrieval
|
| 5 |
-
- reasoning-retrieval
|
| 6 |
- bright
|
| 7 |
-
- information-retrieval
|
| 8 |
language:
|
| 9 |
- en
|
| 10 |
---
|
| 11 |
|
| 12 |
# Darwin-BRIGHT-Retriever
|
| 13 |
|
| 14 |
-
|
| 15 |
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
1. **Query reasoning augmentation** with our Darwin reasoning model: for each query, the system first reasons about what an ideal answer passage must contain (key concepts, theorems, terms, entities), and appends that reasoning to the query.
|
| 19 |
-
2. **Dense retrieval** with the open embedder **Qwen3-Embedding-4B** over the augmented query.
|
| 20 |
-
|
| 21 |
-
The contribution here is the Darwin reasoning-augmentation stage. The embedder is a public, third-party model used as-is; it is named for full transparency and is not claimed as ours.
|
| 22 |
|
| 23 |
## Evaluation
|
| 24 |
|
| 25 |
-
Measured with the MTEB evaluation framework on
|
| 26 |
-
|
| 27 |
-
Baselines reported alongside for context:
|
| 28 |
-
- BM25 (public baseline): reproduced on the same tasks.
|
| 29 |
-
- Qwen3-Embedding-4B alone (public embedder): reference.
|
| 30 |
-
|
| 31 |
-
> Evaluation in progress. Final aggregate score is published only after all tasks complete and are measured end to end. No partial or estimated numbers are reported.
|
| 32 |
-
|
| 33 |
-
## Method notes
|
| 34 |
-
|
| 35 |
-
- Query-only augmentation; documents are embedded unchanged.
|
| 36 |
-
- Darwin generates a short expansion per query (2-3 sentences), then the augmented query is encoded.
|
| 37 |
-
- Deterministic, reproducible pipeline over the released BRIGHT public queries and corpora.
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
tags:
|
| 4 |
- retrieval
|
|
|
|
| 5 |
- bright
|
|
|
|
| 6 |
language:
|
| 7 |
- en
|
| 8 |
---
|
| 9 |
|
| 10 |
# Darwin-BRIGHT-Retriever
|
| 11 |
|
| 12 |
+
FINAL-Bench in-house retrieval system for the **BRIGHT** benchmark (reasoning-intensive information retrieval).
|
| 13 |
|
| 14 |
+
Internally developed. Method not disclosed.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
## Evaluation
|
| 17 |
|
| 18 |
+
Measured with the MTEB evaluation framework on **BRIGHT (v1.1)**, nDCG@10, under the public leaderboard protocol. Scores are published only after the full evaluation completes and is measured end to end.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|