Tasksource-JEV-Nano-v0
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.6/10
Apache-2.0 weights and a clear card with reproducible pylate inference. No /v1/systemone server, no first-party PyPI package, no hosted SLA. Day-old Hugging Face release.
- Capability5.8/10
First-class choice / noul / score with a published learned temperature (0.53) and order-independent MaxSim scoring. No p50/p99 latency; author benches sit mid-pack vs GLiClass/GLiNER peers on the same 1k samples.
- Adoption7/100
Launch day: 2 Hugging Face likes, 0 downloads at check time. The companion typed-decisions dataset and the tasksource org add a thin ecosystem signal; no production reports.
Vendor claims
- Typed Decisions 42.6%; AG News 73.8%; Emotion 49.0%; Banking77 44.7%; Fast Decisions 41.3% exact; classifier-benchmark v2 57.5 micro / 57.2 macro (1k samples each, author protocol)[Vendor claim — not independently verified]
- Tasksource unseen tasks 58.9%; unseen test tasks 53.0% (checkpoint selected on unseen tasks at step 4000/8000)[Vendor claim — not independently verified]
- Outputs are exactly permutation-equivariant; situation encoding can be cached across candidate sets[Vendor claim — not independently verified]
- Fine-tuned on 512k real decisions across 519 tasks (choice / noul / score mix); no teacher model[Vendor claim — not independently verified]
News covering this model
Tasksource-JEV-Nano-v0 is a small open decision model that scores options with multi-vector late interaction. You give it a situation (state + question) and a list of candidates for choice, noul or score; it returns a softmax over MaxSim scores. It does not generate text and it is not a TypeSafe product.
It is a fine-tune of LightOn’s LateOn (~149M) loaded through PyLate / ColBERT-style encoders. The situation is encoded once into token vectors and reused for every candidate set; each option is encoded independently. The release ships a learned temperature of 0.53.
Weights: Hugging Face tasksource/tasksource-jev-nano-v0 (Apache-2.0). Training data: tasksource/tasksource-jev-typed-decisions (HF license tag other — check before commercial use). Author: Damien Sileo. HF repo created .
Specs
| Attribute | Value |
|---|---|
| Author | Damien Sileo / tasksource |
| Base model | lightonai/LateOn (~149M) |
| Readout | ColBERT MaxSim late interaction (not a classifier head) |
| Calibration | Learned temperature 0.53 (card) |
| Decision types | choice, noul (yes/no as two options), score (ordered scale) |
| License | Apache-2.0 (weights); dataset license tagged other on HF |
| Released | |
| Runtime | pip install -U pylate torch — custom encode/MaxSim script on the card |
TypeSafe /v1/systemone |
No |
Latency is not published on the card; the catalog band uses 0 / 0 as an unavailable-data sentinel (same pattern as GLiNER2.5-Decide).
Author benchmarks
All numbers come from the model card. They are author-reported → UnverifiedClaim.
| Benchmark (1k samples, author protocol) | v0 |
|---|---|
| Tasksource unseen tasks | 58.9% |
| Tasksource unseen test tasks | 53.0% |
| Typed Decisions | 42.6% |
| AG News | 73.8% |
| Emotion | 49.0% |
| Banking77 | 44.7% |
| Fast Decisions (2,900 / 17 domains) | 41.3% exact |
| classifier-benchmark v2 | 57.5 micro / 57.2 macro |
The card also places v0 next to compact peers (GLiClass / GLiNER2.5) on the same 1k samples and notes that Laya’s higher AG News figure overlaps its training mix. Treat every peer comparison as UnverifiedClaim until independently rerun.
Limits
- Custom stack. PyLate + MaxSim — not
AutoModelForCausalLM, not a Jev HTTP drop-in. - No latency or ECE table on the public card beyond the learned temperature.
- Dataset license. The companion dataset is tagged
otheron Hugging Face; the Apache-2.0 weight license does not automatically clear training-data provenance for every commercial path. - Day-zero traction. Two likes at first check; APIs and recipes may still move.
- Not TypeSafe. Multi-vector System One–style decisions ≠ affiliated weights or support.
Fit / anti-fit
Fit when you need: a tiny open encoder that caches the situation across many candidate sets, order-independent option scoring, or a research baseline under 200M parameters.
Anti-fit when you need: /v1/systemone wire compatibility, a hosted SLA, published p50/p99, or a number you can treat as frontier-calibrated without reading the card.
Why it is in the catalog
It passes the trained-weights test: LateOn was fine-tuned on hundreds of thousands of real typed decisions, with choice / noul / score in the training mix — not a logit wrapper on a frozen chat model. Weights and an inference recipe are public. It sits near Julia-1, GLiNER2.5-Decide and Laya in the small-encoder band, with a different scoring geometry (late interaction instead of a classifier head).
Derivative: kleos-decider-nano (23 Oct 2026)
publicado (OK do Fabricio 24/10 BRT).
kleos-decider-nano (Mazino0, MIT, created 23 Oct) is a 150M fine-tune of tasksource-decider-nano for agent memory: does a new fact supersede an old one, contradiction / duplicate / compatible, same entity, relation. It is the built-in decider of the kleos database. Author numbers (UnverifiedClaim, kleos’s own Rust eval): supersedes AUROC 0.938 → 0.991 on 66 items; LongMemEval knowledge-update 0.386 → 0.824; MultiNLI dev 0.805 → 0.817 as a forgetting check. A specialist derivative with no board row, so no separate page; scores of this page unchanged.