Etsu-Embedding-0.8B
Name: Etsu โ from the Japanese reading of ไผ้ (etsลซ), the Buddhist term for the convergence of disparate teachings. From Zheng Qiao's (้ๆจต) Tongzhi (ใ้ๅฟยทๆปๅบใ):
็พๅทๅผ่ถ๏ผๅฟ ไผไบๆตท๏ผ็ถๅไนๅทๆ ๆตธๆทซไนๆฃ๏ผไธๅฝๆฎ้๏ผๅฟ ้่ฏธๅค๏ผ็ถๅๅ ซ่ๆ ๅฃ ๆปไนๅฟงใไผ้ไนไนๅคง็ฃๅ๏ผ
A hundred rivers run different courses, yet all must reach the sea; ten thousand states take different roads, yet all converge upon the Central States โ great indeed is the meaning of huitong (convergence). An embedding model does the same: it maps heterogeneous inputs into one shared vector space.
An embedding model fine-tuned from Qwen/Qwen3.5-0.8B-Base for agent memory retrieval and related domain-specific semantic retrieval.
Model Overview
- Model Type: Text Embedding
- Supported Languages: Chinese, English
- Number of Parameters: 752M (752,393,024)
- Context Length: 2048 tokens
- Embedding Dimension: 1024 (supports MRL dimensions 128, 256, 512, 768, 1024)
Model Details
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3.5-0.8B-Base |
| Architecture | Qwen3.5 hybrid text model |
| Layers | 24 (18 linear-attention + 6 full-attention) |
| Hidden size | 1024 |
| Intermediate size | 3584 |
| Vocabulary size | 248,320 |
| Embedding output | 1024 dimensions |
| Pooling | Last-token pooling |
| Normalization | L2 normalization |
| Similarity | Cosine similarity or normalized dot product |
| Training precision | BF16 |
| MRL dimensions | 128, 256, 512, 768, 1024 |
| License | Apache-2.0 |
Intended Use
Chinese/English semantic retrieval for agent memory, personal knowledge bases, conversation archives, technical documentation, and related domain-specific search, including short keyword queries and natural questions.
Encoding Protocol
Queries:
Instruct: Given a query, retrieve relevant passages that answer the query
Query: {query}
Memories/passages: no instruction prefix.
Use last-token pooling and L2 normalization. If using Matryoshka truncation, truncate to 128, 256, 512, or 768 dimensions before the final L2 normalization.
This is not a sentence-transformers package with a built-in pooling wrapper; implement the pooling and query prefix as described above.
Comparison with Qwen/Qwen3-Embedding-0.6B
S2 is this model; Official is Qwen/Qwen3-Embedding-0.6B. Values are Recall@K or MRR. The same tokenizer protocol, query instruction, pooling, normalization, corpus, and query set were used per row.
OpenClaw fine-tuning validation
| Dim | S2 R@1 | Official R@1 | S2 R@3 | Official R@3 | S2 R@5 | Official R@5 | S2 R@10 | Official R@10 | S2 MRR | Official MRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 128 | 8.96% | 17.19% | 19.57% | 31.84% | 25.82% | 37.98% | 35.32% | 46.49% | 17.57% | 27.38% |
| 256 | 15.04% | 20.12% | 30.74% | 35.99% | 38.58% | 43.67% | 50.14% | 52.35% | 26.46% | 31.25% |
| 512 | 21.67% | 22.33% | 40.57% | 39.19% | 49.75% | 47.21% | 62.02% | 57.16% | 34.84% | 34.11% |
| 768 | 23.88% | 22.94% | 44.67% | 40.57% | 53.45% | 48.31% | 65.06% | 58.71% | 37.60% | 34.97% |
| 1024 | 25.21% | 23.66% | 45.55% | 41.46% | 54.62% | 49.36% | 66.56% | 59.92% | 38.86% | 35.66% |
Unseen memory
Corpus of 382 memory chunks dated after 2026-08-08; 1,048 Gemini-generated queries with multi-positive qrels where applicable.
| Dim | S2 R@1 | Official R@1 | S2 R@3 | Official R@3 | S2 R@5 | Official R@5 | S2 R@10 | Official R@10 | S2 MRR | Official MRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 128 | 12.02% | 18.80% | 26.15% | 40.55% | 35.50% | 49.43% | 47.81% | 59.73% | 23.76% | 33.14% |
| 256 | 17.08% | 20.32% | 35.40% | 43.89% | 45.99% | 51.91% | 61.83% | 62.02% | 31.00% | 35.07% |
| 512 | 20.42% | 23.09% | 43.23% | 45.32% | 54.87% | 54.10% | 68.61% | 63.93% | 36.48% | 37.40% |
| 768 | 21.56% | 22.04% | 45.80% | 45.80% | 59.06% | 54.39% | 72.23% | 63.84% | 38.22% | 36.91% |
| 1024 | 22.33% | 21.09% | 46.47% | 45.52% | 58.59% | 53.91% | 72.04% | 64.31% | 38.80% | 36.25% |
Conventional content validation sample
Stratified sample of 382 queries from the pretraining validation split, with 761 unique positive/negative passages.
| Dim | S2 R@1 | Official R@1 | S2 R@3 | Official R@3 | S2 R@5 | Official R@5 | S2 R@10 | Official R@10 | S2 MRR | Official MRR |
|---|---|---|---|---|---|---|---|---|---|---|
| 128 | 80.37% | 81.41% | 93.46% | 93.19% | 96.60% | 96.34% | 98.43% | 97.91% | 87.45% | 87.93% |
| 256 | 87.70% | 85.86% | 96.60% | 93.98% | 98.43% | 96.34% | 99.74% | 97.64% | 92.42% | 90.43% |
| 512 | 89.53% | 85.34% | 97.91% | 93.72% | 99.21% | 96.34% | 99.74% | 98.43% | 93.71% | 90.22% |
| 768 | 91.10% | 84.82% | 97.64% | 95.03% | 99.21% | 97.38% | 99.74% | 98.69% | 94.58% | 90.34% |
| 1024 | 91.36% | 84.29% | 97.64% | 95.29% | 98.95% | 97.12% | 99.74% | 98.69% | 94.74% | 90.13% |
Recommendations
- OpenClaw's default memory search returns up to 10 results, so Recall@10 is the primary metric.
- Use 1024 dimensions for maximum ranking quality, especially when top-1/top-3 and MRR matter.
- Use 768 dimensions as the storage/Recall@5-10 compromise; it uses about 25% less vector storage than 1024 dimensions.
Training Data Disclosure
The model was trained on Chinese retrieval data assembled from two public datasets,
combined, deduplicated and shuffled (ensure_data() in train.py):
| Source | Rows | Fields | License / Origin |
|---|---|---|---|
sentence-transformers/t2ranking (train) |
600,000 | query / positive / negative | HF dataset card |
sentence-transformers/dureader (train) |
180,000 | query / positive / negative | HF dataset card |
all_nli.jsonl (English NLI) |
120,000 | sentence pairs, lang: en |
mixed NLI (see below) |
After dedup + shuffle: train.jsonl 889,836 rows, valid.jsonl 8,988 rows
(900,000 raw โ 898,824 unique; valid = min(10000, len/100)).
File checksums (SHA-256)
b396d0e0a6bc8cfa362a5dbebc75d39fe34957c451071d4fcc4ad577fa6713cf t2ranking.jsonl
2bf403211c0f4249a4f3c795c66c09b7d4633fbccb12e1f7331ae76ab535447a dureader.jsonl
1ec56c99bf9b3d8ff3d9dd373af28ba2c06ffbab9c584b344a4386e57925a71e train.jsonl
a81323db927859ec05bfe23af70e6c99b3faec4063fc8021767bdb9afe59117a valid.jsonl
4ca2c0a6e701d95e0a7b1ea31d1f1c78da99d937fa58fa3471e9a3c023a9ce7a all_nli.jsonl
Benchmark overlap audit (2026-10-17)
Compared against the query/corpus hashes of 8 Chinese retrieval tasks
(C-MTEB 7 + mteb/T2Retrieval), covering 867,992 query+passage strings:
- Query sets: only 3 unique eval queries appear in training (T2Retrieval 2/22,812; DuRetrieval 1/2,000; others 0).
- Passage (corpus) sets, unique eval passages also present in training:
- T2Retrieval: 4,016 / 118,605 (3.39%)
- CovidRetrieval: 984 / 100,001 (0.98%)
- DuRetrieval: 150 / 100,001 (0.15%)
- CmedqaRetrieval: 16 / 100,001 (0.02%)
- MedicalRetrieval: 7 / 100,999 (0.01%)
- MMarcoRetrieval / EcomRetrieval / VideoRetrieval: 0
- Overlapping passages come almost entirely from the
negativefields (t2ranking โ T2Retrieval corpus 3,953; dureader โ CovidRetrieval 823), i.e. the model saw them as random negatives, not as positive targets. Positive-field overlap is only 282 passages in total.
Interpretation: no meaningful queryโanswer leakage; the residual overlap is source-level corpus exposure (same web QA pool), which we disclose here rather than claim a fully clean split. Scores on T2Retrieval/CovidRetrieval should be read with this caveat.
Base Model and License
Derived from Qwen/Qwen3.5-0.8B-Base and distributed under the base model's Apache-2.0 license. Please review the base model license and usage requirements.
- Downloads last month
- 314