Jeff runs zero-shot classification in 22 ms using a 0.8B local model

·5 views

Jeff is an open-source project published on GitHub that performs zero-shot classification using a 0.8-billion-parameter model, delivering results in 22 milliseconds on an RTX PRO 6000 GPU and 28 ms on an Apple M4 Max. The model can assign inputs to categories it never encountered during training by accepting natural-language descriptions of options at query time, returning a calibrated probability per option in a single forward pass. Fine-tuning the base Qwen3.5-0.8B model with Jeff's approach raised its average benchmark score from 45.3 to 79.1, approaching the 83.0 reported by the larger Jev model. Jeff excels at classification tasks — scoring 96.4 on Financial PhraseBank versus Jev's 77.0 — but trails significantly on reasoning benchmarks like BBH, where Jev scores 94.3 against Jeff's 64.0. The entire training pipeline runs on local hardware using synthetic data generated by an open model, with no reliance on cloud GPUs or proprietary model outputs.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Why AI Agents Struggle to Distinguish Fresh Memory from Stale Data

A developer building an AI sales assistant found that giving an agent persistent memory does not solve the context problem — it merely relocates it. The core challenge shifts from fitting all history into a prompt to retrieving only the most relevant and current information for a given query. Outdated or conflicting customer data, such as an objection that was later resolved, can mislead the model if retrieved without temporal context. The system addresses this by logging call outcomes after each interaction, creating a continuous memory loop that updates without retraining the underlying language model. A key design choice was making retrieved memories visible to users, enabling clearer debugging by separating retrieval failures from reasoning failures.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Design Honest Empty States for Features That Are Often Blank

Many UI features — such as live sections, sale banners, or expiring stories — are empty by default, yet most apps display a generic 'nothing here' message that users mistake for a broken page. A better approach distinguishes between four distinct states: expected absence, temporary unavailability, restricted content, and actual errors, since each requires different messaging and a different next step. Developers are advised to return an explicit status field from the API rather than inferring state from an empty array, ensuring the UI always knows why content is missing. When no live content exists, the interface should plainly acknowledge this as normal, explain when content may return if known, and immediately display a useful fallback in the same view. For expiring content, time remaining should be calculated client-side from an absolute timestamp and refreshed every minute to stay accurate even in long-open tabs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Consistent Publishing Hinges on Facts, Not Inspiration, Editors Argue

A content strategy perspective from DEV Community argues that missed publishing deadlines are rarely caused by a lack of creativity, but rather by an absence of concrete facts such as dates, deliverables, and approved claims. When writers lack this factual foundation, they resort to padding content, which digital platforms tend to penalize. The proposed solution is a simple monthly checklist that operators can complete in roughly twenty minutes, covering what shipped, what changed, and what must not be stated. The approach also cautions against common shortcuts like fabricated case studies, stacked calls-to-action, and duplicate posts filed under different titles. The core principle is straightforward: gather the facts first, then begin writing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds ShipCheck, a static preflight scanner to assess Unity project release readiness

A developer has created ShipCheck, a local read-only tool that scans Unity project folders to identify issues that could prevent a game from shipping. The tool separates three distinct concepts — project health, release readiness, and confidence — after early versions conflated all findings into a single misleading score. For older Unity projects where scene serialization cannot be reliably parsed offline, ShipCheck now reports 'LIMITED EVIDENCE' rather than generating an inaccurate readiness number. To filter out irrelevant third-party asset noise, the tool builds a shipping dependency graph starting from enabled build scenes and traces serialized GUID references to classify content as shipping, non-shipping, or third-party. The project also accounts for Unity's runtime content-loading patterns, ensuring assets not directly referenced in scenes are still considered during analysis.

0 comments Read more at DEV Community