Update model card: listwise generative reranker fine-tune on Qdrant hard negatives

README.md

@@ -1,19 +1,76 @@
 ---
 license: apache-2.0
 base_model: Qwen/Qwen3-Reranker-0.6B
+base_model_relation: finetune
+library_name: transformers
+pipeline_tag: text-ranking
+language:
+  - en
+  - vi
 tags:
   - reranker
   - cross-encoder
+  - text-ranking
+  - listwise-reranking
+  - generative-reranker
+  - cve
+  - cybersecurity
+  - qdrant
   - secAI
 ---
 
 # sec-rerank
 
-This is a direct re-host of [Qwen/Qwen3-Reranker-0.6B](https://huggingface.co/Qwen/Qwen3-Reranker-0.6B)
-under the `secAI` project namespace, served as the reranker in the secAI retrieval pipeline
-(via vLLM, pooling runner, sequence-classification mode). Weights are unmodified from the
-base model.
+Domain-adapted **listwise generative reranker** for CVE / cybersecurity search. Fine-tuned from [Qwen/Qwen3-Reranker-0.6B](https://huggingface.co/Qwen/Qwen3-Reranker-0.6B) (Apache-2.0) on grouped query–document examples whose **hard negatives were mined from a local Qdrant** collection (`cve_kb`).
 
-- **Base model:** Qwen/Qwen3-Reranker-0.6B (Apache-2.0 license)
+This is a **second-stage ranker**, not an embedding model. It does not produce dense vectors. It reorders first-stage retrieval candidates (e.g. from [`DuyTa/sec-embedding`](https://huggingface.co/DuyTa/sec-embedding)) by scoring each `(query, document)` pair with the native Qwen3-Reranker generative head.
 
-All credit for the underlying model goes to the original Qwen team.
+## Training
+
+Recipe follows the secAI Colab trainer (`notebooks/Qwen3_Reranker_Colab.ipynb`):
+
+| | |
+|---|---|
+| Base | `Qwen/Qwen3-Reranker-0.6B` |
+| Task | SWIFT `generative_reranker` (causal-LM reranker, not a bi-encoder) |
+| Loss | **listwise reranking** (`--loss_type listwise_reranker`) |
+| Tuner | full-parameter SFT (`--tuner_type full`) |
+| Engine | [ms-swift](https://github.com/modelscope/ms-swift) `swift sft` |
+| Max length | 2048 |
+| Learning rate | 6e-6 |
+
+Project `(query, positive, negative)` rows are converted to SWIFT's **grouped ranking** schema before training:
+
+- `messages` — system instruction + user query
+- `positive_messages` — gold CVE passage
+- `negative_messages` — hard-negative CVE passage(s)
+
+Instruction used at train time:
+
+> Given a Vietnamese cybersecurity search query, retrieve passages from the CVE knowledge base that directly answer it.
+
+SWIFT fills the native Qwen3-Reranker `{Instruction}` slot from that system message. Do not inject `<|im_start|>` or `<think>` into the JSONL.
+
+### Hard-negative mining
+
+Positives and negatives are real CVE core-chunk text from the local Qdrant `cve_kb` (NVD/MITRE). For each query, the negative is a **near-miss CVE** from the same collection — typically a different CWE (`hard_negative_type: different_cwe`) so lexical overlap is high but the relevant document is wrong. That is the listwise signal: rank the gold passage above mined hard negatives for the same query.
+
+Full split is 33.6k train / 4.2k validation grouped examples.
+
+## Inference
+
+Keep the Qwen3-Reranker prompt format. The model scores a candidate by the generative yes/no head; higher score = more relevant. Use it only to **rerank** a short candidate list from dense / hybrid retrieval, not as a first-stage embedder.
+
+```python
+# vLLM / OpenAI-compatible rerank endpoint (secAI serving stack)
+# POST /v1/rerank
+{
+  "model": "DuyTa/sec-rerank",
+  "query": "CVE-2021-44228 JNDI lookup on log4j",
+  "documents": ["...", "..."]
+}
+```
+
+## Attribution & license
+
+Weights derive from [Qwen/Qwen3-Reranker-0.6B](https://huggingface.co/Qwen/Qwen3-Reranker-0.6B), released under **Apache-2.0**. All credit for the base reranker belongs to the Qwen team. This card and the Qdrant listwise fine-tune are part of the secAI retrieval stack.