SKILLRET-Edge-22M-int4

Engineering blog: Skill Retrieval on a CPU: Building the SKILLRET Small Model Family (한국어: CPU에서 도는 스킬 검색)

A 22.7M-parameter bi-encoder for agent skill retrieval — picking the right skill out of a catalogue for a natural-language request. Distilled from ThakiCloud/SKILLRET-Embedding-0.6B and small enough to run on a CPU next to the agent.

Scored on the public ThakiCloud/SKILLRET test split (4,392 queries / 6,006 skills), the metric the SkillRet paper uses as its headline.

NDCG@10 75.14 · 17.0 MB packed

Correction — September 2026. Earlier versions of this card reported the SKILLRET-Embedding-0.6B teacher as 80.82 NDCG@10 on the current 4,392-query / 6,006-skill evaluation split. That score was measured before the train/evaluation query-prefix contract was corrected. Re-evaluation on ThakiCloud/SKILLRET revision a050ad2 under the canonical query contract gives 78.48 NDCG@10. The model weights are unchanged. The earlier 605 MB size was also not the serialized artifact size; the distributed BF16 checkpoint is 1191.6 MB (decimal MB). The "vs teacher" column was also still computed against 80.82 in earlier versions of this card; it is now computed against 78.48, the teacher score measured under the same evaluation condition as the rows below. We retain this note so results quoted from earlier versions of the card can be interpreted correctly. Correction — November 2026: two rows of "What did NOT work" (8B-teacher distillation, hard-negative mining) are now marked as past experiments not re-verified under the current evaluation contract; statements of a statistical tie or advantage that were not backed by a paired test were removed; and where a teacher latency appeared, it is now reported separately from the student latency with its own measurement conditions.

Results

int4 costs 0.12pp against fp16 — a paired t of 1.15, i.e. not a statistically significant difference — while cutting the file 2.7x.

CPU latency

This model, PyTorch, fp32 arithmetic, 4 threads, one short query or a batch of 32:

⚠️ Two different machines, so this is not an isolated ISA comparison — memory and clock differ too. fp32 arithmetic, weights-only quantization.

Separate measurement, not comparable with the table above — the 0.6B teacher (SKILLRET-Embedding-0.6B) exported to ONNX with INT8 weights, Apple M4 Pro, ONNX Runtime with default threads, sequence length 512, 60 timed runs after 5 warm-up runs:

Runtime, precision, thread count and input length all differ between the two measurements, so do not divide one by the other to get a speed-up. Earlier versions of this card put this teacher's p50–p99 range (518–650 ms) in the student table as "same bench"; that label was wrong.

Usage

from sentence_transformers import SentenceTransformer
m = SentenceTransformer("ThakiCloud/SKILLRET-Edge-22M-int4")

# ⛔ Encode queries BARE — no instruction prefix. See query_prefix.json.
q = m.encode(["I need to convert a spreadsheet into a chart"], normalize_embeddings=True)
d = m.encode(["Chart Builder — turn tabular data into bar/line charts"], normalize_embeddings=True)
print(q @ d.T)

What "int4 / g16" means here

Weights are quantized group-wise, asymmetric min/max, group size 16, with scales and zero-points stored in fp16 — the same shape as a GGUF Q4_K/Q8_0 block.

This repo ships two things, deliberately:

Be clear about what this does and does not buy you. The packed file is the real storage footprint and it is verified: predicted 17.074 MB vs 17.012 MB on disk (0.36% error). But the safetensors path runs fp32 arithmetic — the latency figures below were measured that way, and they are not integer-kernel speeds. Routing the packed weights through a real INT kernel (llama.cpp, ONNX Runtime) is unmeasured and is where further speedup would come from.

quantization.json carries the per-tensor layout if you want to write that kernel.

How it was trained

Knowledge distillation from ThakiCloud/SKILLRET-Embedding-0.6B (NDCG@10 78.48 on the same split), kd_weight=0.7, multi-positive InfoNCE over 1–3 golds per query, 12 epochs, cosine schedule. The epoch was chosen on a skill-disjoint holdout — the test split was never used for selection.

⚠️ Read this before you compare numbers

Use the same query prefix at train and eval time. This repo ships query_prefix.json recording the contract (resolved: "", i.e. queries are encoded bare, with no instruction prefix). Which prefix you pick made little difference after fine-tuning in our runs — none 79.18 vs an instruction prefix 78.50 for the 109M student, a 0.68-point gap we did not test for significance — but keeping it consistent matters a great deal. We measured a +8.84pp swing on a single checkpoint purely from a train/eval prefix mismatch, and that mismatch also manufactured a fake result in which quantization appeared to beat fp16. It does not.

Other caveats worth stating plainly:

• The model card of the SkillRet reference models reports a 4,997-query / 6,660-skill split. That split is not in the currently published dataset — we verified the public files are hash-identical to ours at 4,392 / 6,006. Do not convert between the two.
• Standard error on this split is about ±0.45. Differences under ~1pp are not rankings.
• Pooling is CLS, embeddings are L2-normalized, max_length=256 at evaluation.

What did NOT work (so you don't repeat it)

License

Apache-2.0, inherited from the base model.

Citation

Benchmark: SkillRet, arXiv:2605.05726

Paper (measurements behind the quantized variants and the student frontier): Where Post-Training Quantization Breaks Text Embedders: A Measured Map Across Four Embedder Families, arXiv:2609.16391

@article{han2026ptqembedders,
  title   = {Where Post-Training Quantization Breaks Text Embedders: A Measured Map Across Four Embedder Families},
  author  = {Han, Hyojung},
  journal = {arXiv preprint arXiv:2609.16391},
  year    = {2026}
}