Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational Agents

Source: Hugging Face - Blog | Tags: ecommerce, llm, rl, verifiable-environments | Published: 2026-04-16 | Novelty: 35%

EcomRLVE-GYM introduces eight verifiable e-commerce environments for training large language models, including Cart Building, Product Discovery, and Journey. It uses Qwen 3.5 to simulate realistic user interactions and DAPO algorithm for reinforcement learning, demonstrating progressive growth in difficulty reached through adaptive scheduling. Notable is the use of a 2M-product catalog from Amazebay-catalog-2M dataset.

llm-anthropic 0.25

Source: Simon Willison's Weblog | Tags: claude-opus-4.7, models, options, thinking-effort | Published: 2026-04-16 | Novelty: 35%

llm-anthropic 0.25 introduces claude-opus-4.7 with high thinking_effort support and new boolean options for thinking_display and thinking_adaptive. It also increases the default max_tokens to the maximum allowed per model and removes obsolete structured-outputs-2025-11-13 beta header for older models.

Building a Fast Multilingual OCR Model with Synthetic Data

Source: Hugging Face - Blog | Tags: multilingual, ocr, synthetic-data | Published: 2026-04-17 | Novelty: 35%

The article introduces Nemotron OCR v2, a multilingual OCR model that leverages synthetic data generated from mOSCAR text and SynthDoG rendering. The model, trained on 12 million synthetic images across six languages, achieves significant accuracy improvements with Normalized Edit Distance scores down to 0.035-0.069 for non-English languages compared to previous versions. Notably, the architecture unifies detection and recognition into a single network using RegNetX-8GF as the backbone, enabling fast inference at 34.7 pages/second on an A100 GPU.

Designing synthetic datasets for the real world: Mechanism design and reasoning from first principles

Source: The latest research from Google | Tags: evaluation, mechanism-design, reasoning-driven, synthetic-data, taxonomic-coverage | Published: 2026-04-16 | Novelty: 34%

Simula is a reasoning-first framework for generating synthetic datasets that decomposes the generation process into global diversification, local diversification, complexification, and quality checks. It uses taxonomic coverage and calibrated complexity scoring to evaluate dataset diversity and difficulty, demonstrating superior performance across five diverse domains including cybersecurity and legal reasoning. The framework combines global coverage, local diversity, and critiquing to consistently outperform simpler baselines.

datasette 1.0a28

Source: Simon Willison's Weblog | Tags: compatibility, database, datasette, testing | Published: 2026-04-17 | Novelty: 17%

Datasette 1.0a28 addresses compatibility issues introduced in the previous alpha version, including fixing a bug where execute_write_fn() callbacks with parameters other than conn were encountering errors (#2691). Additionally, new features such as datasette.close() for closing all databases and resources associated with an instance have been added. This method is called automatically when the server shuts down, and Datasette now includes a pytest plugin that ensures temporary instances are properly cleaned up during testing.