Model card edits: hero image, nDCG@10 table headers, BEIR results section, decontaminated wording

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* 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
+rerank.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -40,6 +40,10 @@ tags:
 <a href="https://huggingface.co/lightonai/LightOn-rerank-LW-4B">LW-4B</a>
 </p>
 
+<div align="center">
+<img src="rerank.png" alt="Reranking: the second stage that puts the best text and image candidates on top" width="560">
+</div>
+
 ---
 
 ## About the LightOn-rerank family
@@ -50,16 +54,16 @@ The models are built on Qwen3.5 vision-language backbones (hybrid linear + full
 
 The family comes in two scoring flavours × three sizes (0.8B / 2B / 4B):
 
-- **PW (pointwise)** — each candidate is scored independently: the model judges whether the document answers the query, and the score is `logit("Yes") − logit("No")`. One forward pass per candidate, no generation — simple to serve (vLLM-compatible) and embarrassingly parallel.
-- **LW (listwise)** — generative listwise ranking: 4 candidates are placed in a single prompt and the model generates a permutation (`[2] > [4] > [1] > [3]`). Larger candidate pools are ranked with a sliding window (window 4, stride 2, bottom-to-top). Cross-document attention makes LW markedly stronger on hard visual reranking, and unlike pointwise scoring it keeps improving with backbone size.
+- **PW (pointwise)**: each candidate is scored independently. The model judges whether the document answers the query, and the score is `logit("Yes") − logit("No")`. One forward pass per candidate and no generation, so it is simple to serve (vLLM-compatible) and embarrassingly parallel.
+- **LW (listwise)**: generative listwise ranking, where 4 candidates are placed in a single prompt and the model generates a permutation (`[2] > [4] > [1] > [3]`). Larger candidate pools are ranked with a sliding window (window 4, stride 2, bottom-to-top). Cross-document attention makes LW markedly stronger on hard visual reranking, and unlike pointwise scoring it keeps improving with backbone size.
 
-**LightOn-rerank-LW-2B** is the best 2B-class model of the family. Putting 4 candidates in a single forward pass lets documents attend to each other, which fixes the discriminative weakness of independent per-candidate scoring: +0.03 NDCG@10 on ViDoRe V3 over the pointwise sibling with identical backbone and negatives, winning 14 of 16 splits.
+**LightOn-rerank-LW-2B** is the best 2B-class model of the family. Putting 4 candidates in a single forward pass lets documents attend to each other, which fixes the discriminative weakness of independent per-candidate scoring: +0.03 nDCG@10 on ViDoRe V3 over the pointwise sibling with identical backbone and negatives, winning 14 of 16 splits.
 
 ## Results
 
-**ViDoRe V3** (visual document reranking, 8 domains × EN/FR queries), NDCG@10, ColQwen2.5-v0.2 first stage, retrieve 100 / rerank 100. All models — including baselines — were re-evaluated under this same two-stage protocol, so numbers are mutually comparable but not comparable to vendor-reported end-to-end results.
+**ViDoRe V3** (visual document reranking, 8 domains × EN/FR queries), overall nDCG@10, ColQwen2.5-v0.2 first stage, retrieve 100 / rerank 100. All models, including baselines, were re-evaluated under this same two-stage protocol, so numbers are mutually comparable but not comparable to vendor-reported end-to-end results.
 
-| Model | Params | Scoring | ViDoRe V3 overall @10 |
+| Model | Params | Scoring | ViDoRe V3 overall nDCG@10 |
 |---|---|---|---|
 | [LightOn-rerank-LW-4B](https://huggingface.co/lightonai/LightOn-rerank-LW-4B) | 4.5B | listwise | 0.6469 |
 | *Qwen3-VL-Reranker-8B* | *8B* | *pointwise (pooling)* | *0.6423* |
@@ -71,7 +75,7 @@ The family comes in two scoring flavours × three sizes (0.8B / 2B / 4B):
 | [LightOn-rerank-LW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-LW-0.8B) | 0.85B | listwise | 0.5825 |
 | [LightOn-rerank-PW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-PW-0.8B) | 0.85B | pointwise | 0.4820 |
 
-### ViDoRe V3 detail (NDCG, ColQwen2.5 first stage, rerank-100, sliding window 4/2)
+### ViDoRe V3 detail (nDCG, ColQwen2.5 first stage, rerank-100, sliding window 4/2)
 
 | Domain | EN @5 | EN @10 | FR @5 | FR @10 |
 |---|---|---|---|---|
@@ -85,31 +89,30 @@ The family comes in two scoring flavours × three sizes (0.8B / 2B / 4B):
 | physics | 0.4596 | 0.4884 | 0.4706 | 0.4988 |
 | **mean** | **0.6246** | **0.6401** | **0.5949** | **0.6131** |
 
-Overall @10: **0.6266** (EN 0.6401 / FR 0.6131) — best 2B-class result under this protocol, +0.033 over the official Qwen3-VL-Reranker-2B and +0.035 over jina-reranker-m0.
+Overall nDCG@10: **0.6266** (EN 0.6401 / FR 0.6131), the best 2B-class result under this protocol (+0.033 over the official Qwen3-VL-Reranker-2B, +0.035 over jina-reranker-m0). On the easier ViDoRe V2 benchmark (same protocol), mean nDCG@10 is **0.8583**, in the same range as jina-reranker-m0 (0.8596) and above MonoQwen2-VL-v0.1 (0.8454).
 
-### Other benchmarks (same fixed-first-stage protocol)
+### BEIR results (text reranking)
 
-| Benchmark | LightOn-rerank-LW-2B | Best 2B-class baseline |
-|---|---|---|
-| ViDoRe V2, mean NDCG@10 (ColQwen2.5 first stage) | 0.8583 | jina-reranker-m0: 0.8596; MonoQwen2-VL-v0.1: 0.8454 |
-| BEIR 15 datasets, clean mean NDCG@10 (jina-embeddings-v3 first stage) | 0.5433 | jina-reranker-m0: 0.5677; Qwen3-VL-Reranker-2B: 0.5025 |
-| BEIR 15 datasets, clean mean NDCG@10 (BM25 first stage) | 0.5040 | jina-reranker-m0: 0.5397; Qwen3-VL-Reranker-2B: 0.4815 |
+15 datasets, nDCG@10, retrieve 100 / rerank 100, all models re-evaluated under the same protocol. The decontaminated mean excludes NQ and MSMARCO (present in the training data).
 
-Clean mean excludes the two contaminated datasets (NQ, MSMARCO).
+| First stage | LightOn-rerank-LW-2B | jina-reranker-m0 | Qwen3-VL-Reranker-2B |
+|---|---|---|---|
+| jina-embeddings-v3 (decontaminated mean) | 0.5433 | **0.5677** | 0.5025 |
+| BM25 (decontaminated mean) | 0.5040 | **0.5397** | 0.4815 |
 
 ## Model Details
 
-- **Model type:** multimodal cross-encoder reranker — **generative listwise**: 4 candidates per prompt, ranked by generating a permutation; sliding window (4, stride 2) for larger pools
+- **Model type:** multimodal cross-encoder reranker (**generative listwise**: 4 candidates per prompt, ranked by generating a permutation; sliding window (4, stride 2) for larger pools)
 - **Base model:** [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) (Qwen3.5 hybrid linear + full attention VLM)
 - **Parameters:** ≈2.2B (bfloat16, 4.4 GB)
-- **Inputs:** query (text) + candidate document(s) — text passage **or** page image
-- **Fine-tuning:** joint text+vision LoRA (r=32, α=32, rsLoRA — merged into the released weights), mixed-modality batches (2 text + 2 vision groups per micro-batch), vision loss weight 1.3, lr 5e-5, warmup 30%, 1 epoch (419 steps)
+- **Inputs:** query (text) + candidate document(s): text passage **or** page image
+- **Fine-tuning:** joint text+vision LoRA (r=32, α=32, rsLoRA, merged into the released weights), mixed-modality batches (2 text + 2 vision groups per micro-batch), vision loss weight 1.3, lr 5e-5, warmup 30%, 1 epoch (419 steps)
 - **Data:** gold ranking permutations built from labelled pos/hard-neg structure (no teacher): 107k text groups (NQ, TriviaQA, MSMARCO) + 106k visual query–page groups (ColPali-family hard negatives at ranks [0, 5, 10])
 - **Languages:** English (training), French (zero-shot transfer)
 - **Requirements:** `transformers >= 5.4.0` (`qwen3_5` architecture)
 - **Internal experiment ID:** `exp34full`
 
-## Usage — generative listwise reranking
+## Usage: generative listwise reranking
 
 The model ranks **4 candidates per prompt** by generating a permutation string such as `[2] > [1] > [4] > [3]`. Candidate pools larger than 4 are ranked with a sliding window (window 4, stride 2) moving from the bottom of the list to the top, so the best candidates bubble up to the front. If a generation cannot be parsed, fall back to the input order (observed fallback rate in our evals: ≈0.01%).
 
@@ -122,7 +125,7 @@ model_id = "lightonai/LightOn-rerank-LW-2B"
 model = AutoModelForImageTextToText.from_pretrained(
     model_id,
     dtype=torch.bfloat16,
-    attn_implementation="flash_attention_2",  # optional — remove if flash-attn is not installed
+    attn_implementation="flash_attention_2",  # optional, remove if flash-attn is not installed
     device_map="cuda",
 ).eval()
 processor = AutoProcessor.from_pretrained(model_id)
@@ -197,12 +200,11 @@ If a window has fewer than 4 candidates, pad it by repeating the last candidate
 
 - Cross-document attention is the key architectural ingredient: a listwise loss over *independently scored* documents regressed vision performance in our ablations, while generative listwise (all 4 documents in one prompt) improves it.
 - Training data is English-only. French works zero-shot (the backbone is multilingual) but is slightly behind English on average.
-- Vision hard negatives were mined with ColPali-family retrievers; the model pairs best with a ColPali-family first stage (e.g. ColQwen2.5) for visual reranking.
-- BEIR contamination flag: NQ and MSMARCO are part of the text training data; headline text figures use clean means that exclude them.
+- BEIR contamination flag: NQ and MSMARCO are part of the text training data; headline text figures use decontaminated means that exclude them.
 
 ## The LightOn-rerank family
 
-| Model | Backbone | Scoring | ViDoRe V3 @10 |
+| Model | Backbone | Scoring | ViDoRe V3 overall nDCG@10 |
 |---|---|---|---|
 | [LightOn-rerank-PW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-PW-0.8B) | Qwen3.5-0.8B | pointwise | 0.4820 |
 | [LightOn-rerank-LW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-LW-0.8B) | Qwen3.5-0.8B | listwise | 0.5825 |