Simplify model training-data reference

README.md

@@ -16,7 +16,7 @@ document tokens.
 
 Training has two stages:
 
-1. supervised hard-negative fine-tuning using the BiCA-prepared training set;
+1. supervised hard-negative fine-tuning;
 2. mixed listwise knowledge distillation on the seven-source hard-negative mixture, with
    `jinaai/jina-reranker-v3.5` scores, temperature sharpening, false-negative masking, and an
    InfoNCE anchor.
@@ -54,12 +54,16 @@ computed by summing, over query tokens, the maximum similarity to a document tok
 | [LateOn](https://huggingface.co/lightonai/LateOn) | 57.22 | 149 | 128 | 50.52 | 47.36 | **39.67** | 45.99 | **92.02** | **53.12** | 79.98 | 45.67 | 37.79 | 63.91 | **89.67** | **21.90** | 76.61 | 83.60 | 30.52 |
 | **GLINT-base** | **57.43** | 149 | 128 | 52.38 | 46.49 | 34.17 | 47.68 | 92.45 | 50.85 | **82.54** | 46.38 | 37.51 | **68.03** | **90.08** | 20.65 | **77.13** | 84.78 | 30.26 |
 
+GLINT-base was evaluated with the project BEIR protocol: corpus IDs are excluded from their own
+query results for ArguAna and Quora, and ArguAna uses 64 query tokens. The 57.43 average is one
+run; it should not be interpreted as a seed-variance estimate.
+
 ## Training data and reproducibility
 
 The corresponding private training artifacts are in
-[GLINT-data](https://huggingface.co/datasets/chungimungi/GLINT-data): the prepared BiCA SFT rows,
-the 1,046,009-row seven-source KD mixture, and Jina teacher-score parquet shards. The repository
-contains no BEIR evaluation corpus or evaluation labels.
+[GLINT-data](https://huggingface.co/datasets/chungimungi/GLINT-data). It contains the complete
+prepared SFT data, the 1,046,009-row seven-source KD mixture, and Jina teacher-score parquet
+shards. The repository contains no BEIR evaluation corpus or evaluation labels.
 
 ## Limitations