SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Many trending model-hub entries are derivatives of a single base model, not new releases

0
·2 views

A developer argues that the trending board on a model hub reflects which weights people are converting, merging and fine-tuning, not which models are best. Many entries trace back to the same base model, such as a widely derived Qwen3.8-27B, appearing as quantizations, merges and LoRA variants. Fine-tunes often reshape output style and formatting rather than add capability, so benchmark gains should be compared against the base on the same harness. The author cites a costly case where an undocumented system preamble baked into a chat template caused two days of misdiagnosed failures. They recommend checking base_model and chat_template fields before trusting evaluation numbers.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

CSS che adatta la galleria al numero di elementi, senza JavaScript

Un articolo su DEV Community propone di gestire il layout di una galleria tramite CSS anziché con props, classi modificatrici o logica JavaScript. L'approccio usa una custom property per il numero di colonne, con un valore predefinito di tre. I selettori :has(), :nth-child() ed :last-child permettono di rilevare se il numero di elementi è pari o superiore a una certa soglia. Con questa regola, le gallerie con due o quattro elementi passano a due colonne, quelle con sette o più a quattro, mentre sei elementi restano su tre. Secondo l'autore, il vantaggio è un componente più pulito che reagisce automaticamente a filtri, paginazione e contenuti dinamici.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

A Valid Payment Signature Does Not Guarantee the Purchase Is Correct

A valid signature only confirms that a payment notification originated from the payment provider, not what the payment should unlock. Developers are advised to test whether changing checkout amounts, product IDs or account references in a sandbox alters what gets granted. The server should control pricing and tie each checkout to the authenticated buyer. Before access is granted, payment status, currency and product should be matched against the stored order, allowing for legitimate discounts and taxes. Test-mode payments must also be kept strictly separate so they can never unlock live access, and custom fields in signed notifications should still be treated as untrusted since they may have originated in the browser.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

osFoundry MCP tool limits GitHub PRs to one dedicated branch push, with owner approval required

The osFoundry MCP server exposes a propose_gitnest_provider_publication tool that lets Codex CLI ask a repository owner to publish a change set to GitHub or GitLab. The tool accepts only a change set ID, version, composition hash, and idempotency key, with branch name, title, and draft settings filled in by the owner at approval time. The author stresses that the approved action pushes only the dedicated branch, and the tool is annotated as open-world and destructive so clients pause for confirmation. A schema test of twelve cases showed limits such as a 20-file cap and a 100,000-character per-file cap, but the 200 KB total is measured in UTF-8 bytes and cannot be enforced by schema alone. The post also notes that the four publication-related tools consume under 1,000 of roughly 12,600 tokens in the full tool list.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Google Launches Gemini Decision 1, a Model That Scores Many Questions in One API Call

Google DeepMind announced Gemini Decision 1 on November 11, 2026, a model designed to produce scores and probabilities rather than generated text. A single request can include a shared multimodal input and up to 1,000 independent yes/no, multiple-choice, or rubric-based questions. Each answer comes with a confidence score, allowing developers to route uncertain results to fallbacks or humans. The shared input is billed once per request, and the endpoint is available now in the Gemini API and Google AI Studio, with integrations rolling out to several third-party frameworks. The source author notes they had not yet tested the model hands-on, so the report is based on the official documentation.

0 comments Read more at DEV Community