--- license: apache-2.0 language: - en - fr base_model: - Qwen/Qwen3.5-2B pipeline_tag: image-text-to-text library_name: transformers tags: - reranker - cross-encoder - multimodal - text-ranking - document-reranking - vidore - beir datasets: - vidore/colpali_train_set - lightonai/embeddings-fine-tuning ---
Reranking: the second stage that puts the best text and image candidates on top [![Website](https://img.shields.io/badge/LightOn-Website-blue?logo=google-chrome)](https://lighton.ai) [![LinkedIn](https://img.shields.io/badge/LightOn-LinkedIn-0A66C2?logo=linkedin)](https://www.linkedin.com/company/lighton/) [![X](https://img.shields.io/badge/@LightOnIO-X-black?logo=x)](https://x.com/LightOnIO) 📝 [Blog post](https://huggingface.co/blog/lightonai/lighton-rerank)

LightOn-rerank-LW-2B

Unified Text + Visual Document Reranker by LightOn

PW-0.8B | LW-0.8B | PW-2B | LW-2B | PW-4B | LW-4B

--- ## About the LightOn-rerank family Production retrieval pipelines usually need two rerankers: one for text passages and one for visual documents (PDF pages, slides, scans). **LightOn-rerank** models are *unified* cross-encoder rerankers: a single model scores both text passages and document page images against a query, on top of any first-stage retriever (BM25, dense embeddings, or ColPali-family late-interaction models). The models are built on Qwen3.5 backbone (hybrid linear + full attention) and jointly fine-tuned on text and visual reranking data with mixed-modality batches (LoRA, merged into the released weights). Training data is English-only; French performance transfers zero-shot from the multilingual backbone. 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 and no generation. - **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: 62.66 nDCG@10 on ViDoRe V3, improving over the ColQwen2.5 first stage by 7.1 nDCG points and leading every open 2B-class multimodal reranker we measured. Putting 4 candidates in a single forward pass lets documents attend to each other, which fixes the discriminative weakness of independent per-candidate scoring: +2.8 nDCG@10 over the pointwise sibling with identical backbone and data, winning 13 of 16 splits. ## 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 nDCG@10 | |---|---|---|---| | [LightOn-rerank-LW-4B](https://huggingface.co/lightonai/LightOn-rerank-LW-4B) | 4.5B | listwise | 64.69 | | *Qwen3-VL-Reranker-8B* | *8B* | *pointwise (pooling)* | *64.23* | | **LightOn-rerank-LW-2B (this model)** | 2.2B | listwise | **62.66** | | [LightOn-rerank-PW-2B](https://huggingface.co/lightonai/LightOn-rerank-PW-2B) | 2.2B | pointwise | 59.87 | | [LightOn-rerank-PW-4B](https://huggingface.co/lightonai/LightOn-rerank-PW-4B) | 4.5B | pointwise | 59.80 | | *jina-reranker-m0* | *2.4B* | *pointwise* | *59.40* | | *Qwen3-VL-Reranker-2B* | *2B* | *pointwise (pooling)* | *59.18* | | [LightOn-rerank-LW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-LW-0.8B) | 0.85B | listwise | 58.25 | | *MonoQwen2-VL-v0.1* | *2B* | *pointwise* | *57.76* | | *First-stage only (ColQwen2.5, no rerank)* | — | — | *55.60* | | [LightOn-rerank-PW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-PW-0.8B) | 0.85B | pointwise | 48.20 | ### ViDoRe V3 detail (nDCG@10, ColQwen2.5 first stage, rerank-100, sliding window 4/2) | Domain | EN | FR | |---|---|---| | finance_en | 71.43 | 62.02 | | finance_fr | 47.59 | 51.08 | | computer_science | 82.26 | 79.23 | | hr | 68.75 | 61.95 | | energy | 67.11 | 70.67 | | industrial | 58.03 | 50.27 | | pharmaceuticals | 68.10 | 65.36 | | physics | 48.84 | 49.88 | | **mean** | **64.01** | **61.31** | Overall nDCG@10: **62.66** (EN 64.01 / FR 61.31), the best 2B-class result under this protocol: +3.5 over the official Qwen3-VL-Reranker-2B (59.18) and +3.3 over jina-reranker-m0 (59.40). The training data is English-only, so the FR column is pure zero-shot cross-lingual transfer; FR even beats EN on three domains (finance_fr, energy, physics). ### BEIR results (text reranking) 13 datasets, nDCG@10, BM25 first stage, retrieve 100 / rerank 100, all models re-evaluated under the same protocol (this model runs the same W=4 stride=2 sliding window it uses for document pages; baselines score through their native text paths). ⚠️ marks datasets in the text training mix (NQ, MSMARCO); the clean mean excludes them. | Dataset | LightOn-rerank-LW-2B | Qwen3-VL-Reranker-2B | jina-reranker-m0 | |---|---|---|---| | fever | **78.92** | 73.34 | 80.51 | | scifact | 75.94 | 77.33 | 79.68 | | trec-covid | 71.78 | **83.47** | 83.25 | | hotpotqa | 71.92 | 75.00 | 76.09 | | nq ⚠️ | 54.84 | 57.44 | 60.51 | | dbpedia | 39.48 | 39.81 | 45.66 | | arguana | **41.25** | 36.78 | 40.68 | | fiqa | 37.54 | 37.50 | 44.64 | | msmarco ⚠️ | 36.95 | 38.44 | 38.21 | | nfcorpus | 35.14 | 37.99 | 37.35 | | touche-2020 | 34.83 | 33.75 | 36.09 | | climate-fever | 23.81 | 25.21 | 26.61 | | scidocs | 18.76 | 19.69 | 21.11 | | **Mean (13)** | 47.78 | 48.90 | **51.57** | | **Clean mean (11, excl. ⚠️)** | 48.12 | 49.08 | **51.97** | The ordering flips relative to ViDoRe V3: on vision this model leads the 2B class; on text jina-reranker-m0 leads decisively, and this model is roughly tied with Qwen3-VL-Reranker-2B. On text the pointwise sibling is slightly stronger ([PW-2B](https://huggingface.co/lightonai/LightOn-rerank-PW-2B): 49.13 clean mean) — the mirror image of vision, where listwise wins at every size. ## 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) - **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), training images resized to 512×512 - **Data:** 213k listwise groups — 107k text groups (NQ, TriviaQA, MS MARCO; each a `[pos, neg_0, neg_1, neg_2]` 4-list with hard negatives mined via the NV-Retriever approach with GTE-ModernBERT) + 106k vision groups (ColPali train set with negatives mined by Nomic). The gold permutation (`pos > neg_0 > neg_1 > neg_2`) is constructed directly from the mining metadata; training is cross-entropy on the permutation tokens only. - **Languages:** English (training), French (zero-shot transfer) - **Requirements:** `transformers >= 5.4.0` (`qwen3_5` architecture) ## 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%). ```python import re import torch from transformers import AutoModelForImageTextToText, AutoProcessor 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 device_map="cuda", ).eval() processor = AutoProcessor.from_pretrained(model_id) PROMPT = "<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n" PERM_RE = re.compile(r"\[(\d)\]\s*>\s*\[(\d)\]\s*>\s*\[(\d)\]\s*>\s*\[(\d)\]") def rank_window(query: str, docs: list[str]) -> list[int]: """Rank exactly 4 text passages; returns window indices, most relevant first.""" body = "\n".join(f"[{i + 1}]: {d}" for i, d in enumerate(docs)) user = f"Query: {query}\n\nRank these passages from most to least relevant:\n{body}\n\nRanking:" inputs = processor(text=[PROMPT.format(user=user)], return_tensors="pt").to(model.device) out = model.generate( **inputs, max_new_tokens=30, do_sample=False, pad_token_id=processor.tokenizer.eos_token_id, ) completion = processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True) m = PERM_RE.search(completion) return [int(g) - 1 for g in m.groups()] if m else list(range(4)) def rerank(query: str, docs: list, window: int = 4, stride: int = 2) -> list[int]: """Sliding-window listwise rerank; returns document indices, most relevant first.""" order = list(range(len(docs))) positions = list(range(max(0, len(docs) - window), -1, -stride)) if positions and positions[-1] != 0: positions.append(0) for pos in positions: end = min(pos + window, len(docs)) p = max(0, end - window) if end - p < 2: continue perm = rank_window(query, [docs[i] for i in order[p:end]]) order[p:end] = [order[p + j] for j in perm] return order query = "What is late interaction in neural information retrieval?" documents = ["passage 1 ...", "passage 2 ...", "passage 3 ...", "passage 4 ...", "passage 5 ..."] print(rerank(query, documents)) ``` For **page images**, build the window prompt with interleaved image placeholders instead: ```python def rank_window_images(query: str, images: list) -> list[int]: # 4 PIL images content = [{"type": "text", "text": f"Query: {query}\n\nRank these documents from most to least relevant:\n"}] for i, img in enumerate(images): content += [ {"type": "text", "text": f"[{i + 1}]: "}, {"type": "image", "image": img}, {"type": "text", "text": "\n"}, ] content.append({"type": "text", "text": "\nRanking:"}) text = processor.apply_chat_template( [{"role": "user", "content": content}], tokenize=False, add_generation_prompt=True ) inputs = processor(text=[text], images=list(images), return_tensors="pt").to(model.device) out = model.generate( **inputs, max_new_tokens=30, do_sample=False, pad_token_id=processor.tokenizer.eos_token_id, ) completion = processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True) m = PERM_RE.search(completion) return [int(g) - 1 for g in m.groups()] if m else list(range(4)) ``` If a window has fewer than 4 candidates, pad it by repeating the last candidate and drop the duplicates from the returned order. Our evaluations ran this model with HF `generate()`. ## Notes & limitations - Cross-document attention is the key architectural ingredient: a listwise *loss* over independently scored documents regressed vision performance well below the pointwise baseline in our ablations; generative listwise (all 4 documents in one prompt) is what flips the sign. For the same reason, don't read this model pointwise at deploy time (e.g. a Yes/No readout) — on our smoke split that cost roughly three quarters of the rerank lift. - **Serve at high image resolution.** The model was trained at 512×512, but inference quality keeps improving with resolution: 62.66 at the processor's default (~14.6k ViT patches/page) vs 61.90 at 1M pixels (−0.8) and 57.06 at the 512² training resolution (−5.6). 1M pixels is a good operating point if ViT compute is a concern. - **To cut latency, shrink the candidate pool, not the scoring method**: reranking the top-20 candidates instead of 100 keeps 85% of the rerank lift at ~5× fewer windows (top-10: 58% at ~12× fewer), measured over the full benchmark. - Training data is English-only. French works zero-shot (the backbone is multilingual) but is slightly behind English on average. - BEIR contamination flag: NQ and MSMARCO are part of the text training data; headline text figures use clean means that exclude them. ## The LightOn-rerank family | 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 | 48.20 | | [LightOn-rerank-LW-0.8B](https://huggingface.co/lightonai/LightOn-rerank-LW-0.8B) | Qwen3.5-0.8B | listwise | 58.25 | | [LightOn-rerank-PW-2B](https://huggingface.co/lightonai/LightOn-rerank-PW-2B) | Qwen3.5-2B | pointwise | 59.87 | | [LightOn-rerank-LW-2B](https://huggingface.co/lightonai/LightOn-rerank-LW-2B) | Qwen3.5-2B | listwise | 62.66 | | [LightOn-rerank-PW-4B](https://huggingface.co/lightonai/LightOn-rerank-PW-4B) | Qwen3.5-4B | pointwise | 59.80 | | [LightOn-rerank-LW-4B](https://huggingface.co/lightonai/LightOn-rerank-LW-4B) | Qwen3.5-4B | listwise | **64.69** | Rule of thumb: **LW** models are stronger at every size (and the gap grows with size); **PW** models are cheaper to serve and score candidates independently. For the best quality pick [LW-4B](https://huggingface.co/lightonai/LightOn-rerank-LW-4B); for the best quality/cost trade-off pick [LW-2B](https://huggingface.co/lightonai/LightOn-rerank-LW-2B); for maximum throughput on text-heavy workloads pick a PW model. ## Citation ``` @misc{ananya2026lightonrerank,   title={One Adapter, Both Modalities: Field Notes from Building and Serving a Multimodal Reranker},   author={Ananya, Ishrat Jahan and Chatelain, Amelie},   year={2026},   howpublished={\url{https://huggingface.co/blog/lightonai/lighton-rerank}}, } ```