Hacker News Reader

Still 10$/month with this referral code?!

Translationauton What happens when an LLM never sees material beyond fifth grade?

There is the art of saying no: https://dl.acm.org/doi/10.5555/3737916.3739489

It is possible to create (subjective) reasoning traces like https://huggingface.co/datasets/Bachstelze/ethical_coconot_6...

And train or adapt a model to it: https://huggingface.co/Bachstelze/olmo-7b-ethical-reasoning-...

This is just a little proof of concept, though it is maybe the direction you are looking for?!

Translationauton Side-by-side comparison of how AI models answer moral dilemmas

The idea of the ethical reasoning dataset is not to erase specific content. It is designed to present additional thinking traces with an ethical grounding. So far, it is only a fraction of the available data. This doesn't solve alignment, and unethical behaviour is still possible, but the model gets a profound ethical reasoning base.

Translationauton Side-by-side comparison of how AI models answer moral dilemmas

There is this ethical reasoning dataset to teach models stable and predictable values: https://huggingface.co/datasets/Bachstelze/ethical_coconot_6... An Olmo-3-7B-Think model is adapted with it. In theory, it should yield better alignment. Yet the empirical evaluation is still a work in progress.

Translationauton Generated Checklists Improve LLM Evaluation and Generation

The authors propose a novel approach where checklists are automatically generated to systematically assess and guide LLM outputs, ensuring more comprehensive and reliable evaluations by LLMs. E.g. it increases in the frequency of exact agreements between LLM judgements and human preferences from 46.4% to 52.2%.

From my perspective it would be neat if the benchmarks would support more model types and not only the predominat GPTs, which only showed that they can relatively easy be scaled up, though it was never stated that they can model language better with the same resources (AFAIK).

Translationauton Show HN: You need a few lines of JS, not a vector database

Why do we need vectors for search anyway? The results are often unrelated to the query. Aren't therefore exact matches better? One could also annotate the corpus with related tags and hypothetical questions, if we need more results.

Translationauton Gemma, Ollama and LangChainGo

E.g. adapters, inference optimization and more (multilingual) models like https://huggingface.co/CohereForAI/aya-101

Translationauton Gemma, Ollama and LangChainGo

What is the advantage of Ollama Python over huggingface?

Translationauton Gemma, Ollama and LangChainGo

What is the point in using Ollama over huggingface if you use Python? Also, REST endpoints can be provided with huggingface transformer.

Here with Go, it seems to make sense to use an abstraction. Though don't you lose a lot of flexibility?

Translationauton StreamingLLM: tiny tweak to KV LRU improves long conversations

This seems only to work cause large GPTs have redundant, undercomplex attentions. See this issue in BertViz about attention in Llama: https://github.com/jessevig/bertviz/issues/128

Translationauton Show HN: Explore large language models with 512MB of RAM

Those minified models are still equal or bigger compared to the initial "attention is all you need" transformer.

Translationauton Show HN: Explore large language models with 512MB of RAM

Have you also tried the bigger models? The smaller models are good for assisted generation: https://huggingface.co/blog/assisted-generation

Those models of LaMini-Flan-T5 are trained to follow instructions and not to recognize the truth content. You could train a transformer like Ernie or Vega (which lead superglue) on such challenging factual data. But don't expect mathematical correct results only from the model. Therefore you have langchain with other APIs.