Desearch Embedding 8B: LoRA adapter on Qwen3-Embedding-8B, model card with /search A/B results

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip 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

1_Pooling/config.json

@@ -0,0 +1,5 @@
+{
+    "embedding_dimension": 4096,
+    "pooling_mode": "lasttoken",
+    "include_prompt": true
+}

README.md

@@ -0,0 +1,102 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-Embedding-8B
+library_name: sentence-transformers
+pipeline_tag: sentence-similarity
+tags: [sentence-transformers, embeddings, retrieval, news, search]
+---
+
+# Desearch Embedding 8B
+
+A LoRA fine-tune of [Qwen/Qwen3-Embedding-8B](https://huggingface.co/Qwen/Qwen3-Embedding-8B) for web and news search, built by
+Desearch. Shipped as a LoRA adapter on the base model: a bf16 merge keeps only 27% of the tuned change, so the change stays in its own fp32 path.
+
+## Usage
+
+```python
+from sentence_transformers import SentenceTransformer
+
+model = SentenceTransformer("desearch/Desearch-Embedding-8B")
+queries = model.encode(["who won the 2026 Tour de France"], prompt_name="query")
+documents = model.encode(["Passage text"])
+scores = model.similarity(queries, documents)
+```
+
+Queries take the instruction prompt `Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:`; documents take none, exactly as the base
+model. A document can be a paragraph chunk, a page head (title and opening text) or a whole page
+up to 2,048 tokens; training covered all three.
+
+## Training
+
+Desearch Embedding 8B adds two LoRA adapters (rank 32 on every attention and MLP projection) to
+Qwen/Qwen3-Embedding-8B, one trained on news and one on reference pages and Wikipedia, then fuses them
+into a single adapter by a weighted sum chosen on held-out questions.
+
+**Documents.** 16,000 documents: news articles from 69 outlets, Wikipedia pages and
+reference pages from the open web, plus Wikipedia pages linked from them for multi-hop questions.
+Every article from the news evaluation week, and every gold page of the public benchmarks below,
+was excluded before sampling.
+
+**Questions.** An LLM wrote 98,307 search questions in the styles people use: keywords,
+natural questions, details from deep in a page, questions with a typo, the central fact of a news
+story with its date, single-answer factual questions, and questions that need two linked pages. A
+second LLM pass graded every candidate passage as answering the question, on topic, or not
+relevant. Only passages graded not relevant became negatives, and every passage graded as
+answering became a positive.
+
+**What the model learns from.** Each positive is trained in the three forms a search index embeds:
+a paragraph chunk, the page head (title and first 1,500 characters) and the full page (up to 2,048
+tokens). Training uses a contrastive loss (temperature 0.02) with in-batch negatives shared across
+GPUs and up to two graded hard negatives per row, and masks any candidate that scores well above
+the positive, so a relevant page that was never graded is not pushed away.
+
+
+## Evaluation
+
+Every comparison is paired, question by question, against the base model under identical settings,
+with an exact sign test. Numbers below are from the held-out half of each question set; the other
+half chose the adapter strength (scale 1.0).
+
+Gold-URL retrieval on a production `/search` index of 44,779 pages, embedded once with Qwen3-Embedding-8B and once with Desearch Embedding 8B; every question is scored by the unchanged desearch-search-evals harness. Cells read base → tuned; a question counts at gold@k when a reference page is in the top k.
+
+**Fast mode, public benchmarks**
+
+| fast | n | gold@1 | gold@5 | gold@10 | recall@10 | wins / losses at gold@1 |
+|---|---|---|---|---|---|---|
+| All | 766 | 85.6% → 86.0% | 96.0% → 96.0% | 96.7% → 97.4% | 82.3% → 83.0% | 15 / 12 (p 0.701) |
+| FRAMES | 399 | 88.7% → 88.7% | 98.5% → 98.2% | 98.7% → 99.0% | 82.7% → 83.0% | 10 / 10 (p 1.000) |
+| SealQA | 137 | 70.8% → 71.5% | 84.7% → 86.1% | 88.3% → 91.2% | 88.3% → 90.3% | 3 / 2 (p 1.000) |
+| SimpleQA-Verified | 230 | 89.1% → 90.0% | 98.3% → 97.8% | 98.3% → 98.3% | 78.2% → 78.6% | 2 / 0 (p 0.500) |
+
+**Fast mode, news**
+
+| fast | n | gold@1 | gold@5 | gold@10 | recall@10 | wins / losses at gold@1 |
+|---|---|---|---|---|---|---|
+| All | 246 | 78.0% → 77.6% | 95.1% → 95.9% | 97.2% → 97.6% | 94.2% → 94.0% | 5 / 6 (p 1.000) |
+
+**Balanced mode, public benchmarks**
+
+| balanced | n | gold@1 | gold@5 | gold@10 | recall@10 | wins / losses at gold@1 |
+|---|---|---|---|---|---|---|
+| All | 766 | 87.7% → 87.9% | 96.6% → 96.3% | 97.4% → 97.1% | 84.2% → 83.9% | 19 / 18 (p 1.000) |
+| FRAMES | 399 | 91.0% → 89.7% | 99.2% → 98.2% | 99.5% → 99.0% | 86.2% → 85.3% | 10 / 15 (p 0.424) |
+| SealQA | 137 | 73.0% → 76.6% | 86.1% → 88.3% | 89.8% → 89.8% | 88.9% → 89.4% | 6 / 1 (p 0.125) |
+| SimpleQA-Verified | 230 | 90.9% → 91.3% | 98.3% → 97.8% | 98.3% → 98.3% | 77.7% → 78.4% | 3 / 2 (p 1.000) |
+
+**Balanced mode, news**
+
+| balanced | n | gold@1 | gold@5 | gold@10 | recall@10 | wins / losses at gold@1 |
+|---|---|---|---|---|---|---|
+| All | 246 | 77.6% → 80.5% | 95.5% → 95.9% | 98.0% → 98.0% | 95.1% → 94.9% | 12 / 5 (p 0.143) |
+
+
+## Limitations
+
+- English only; trained and evaluated on English pages and questions.
+- Queries need the instruction prompt; documents are encoded without one, as with the base model.
+- The reported gains come from one search stack (dense retrieval fused with keyword search) and
+  from the question sets described above. Other corpora and pipelines may see different results.
+- Load it as an adapter on Qwen/Qwen3-Embedding-8B. Merging the change into bf16 base weights keeps
+  only about 27% of it; merge in fp32 if a single set of weights is required.
+- Pages are trained up to 2,048 tokens; longer pages should be split before encoding.
+

adapter_config.json

@@ -0,0 +1,51 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-Embedding-8B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "up_proj",
+    "down_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",
+    "gate_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "FEATURE_EXTRACTION",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e90337b12edca2257dcc386318e6740e3caeec949e92398eb131a16f9a678b20
+size 698416704

chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

config_sentence_transformers.json

@@ -0,0 +1,14 @@
+{
+  "__version__": {
+    "pytorch": "2.13.0+cu130",
+    "sentence_transformers": "5.7.0",
+    "transformers": "5.15.0"
+  },
+  "default_prompt_name": null,
+  "model_type": "SentenceTransformer",
+  "prompts": {
+    "document": "",
+    "query": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
+  },
+  "similarity_fn_name": "cosine"
+}

modules.json

@@ -0,0 +1,20 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.base.modules.transformer.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.sentence_transformer.modules.pooling.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.sentence_transformer.modules.normalize.Normalize"
+  }
+]

sentence_bert_config.json

@@ -0,0 +1,10 @@
+{
+    "transformer_task": "feature-extraction",
+    "modality_config": {
+        "text": {
+            "method": "forward",
+            "method_output_name": "last_hidden_state"
+        }
+    },
+    "module_output_name": "token_embeddings"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff800fee44eacd1a1c8c2f96b890ade65d722550ab6693efa8d2c368f40a5853
+size 11423209

tokenizer_config.json

@@ -0,0 +1,16 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 2048,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "left",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}