Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis9
arxiv.orgauthority6,639 words
phi-1.5 is a 1.3 billion parameter model trained on 30 billion tokens that achieves common sense reasoning benchmark results comparable to models ten times its size. The model exhibits traits of larger LLMs, including the ability to "think step by step" or perform some rudimentary in-context learning, but also shows hallucinations and potential for toxic generations. A variant named phi-1.5-web includes filtered web data, while the base model remains open-source without instruction fine-tuning or any other stage of alignment.written, reviewed, cited
- The model uses a tokenizer from codegen-mono and flash-attention for training speed up.
- None of the models have undergone instruction finetuning or RLHF, yet they can be prompted to follow instructions.
ar5iv.labs.arxiv.orgauthority4,553 words
phi-1.5 is a 1.3 billion parameter model trained on 30 billion tokens that achieves common sense reasoning benchmark results comparable to models ten times its size. The model surpasses most non-frontier LLMs on more complex reasoning tasks such as grade-school mathematics and basic coding. phi-1.5 exhibits traits of larger LLMs including the ability to think step by step but also shows hallucinations and potential for toxic generations.written, reviewed, cited
- phi-1.5 uses a Transformer architecture with 24 layers, 32 heads, and a head dimension of 64.
- The model is trained using a constant learning rate of 2e-4 with no warm-up period.
huggingface.co2023-09-12authorityHN 1 pts · 0 comments397 words
phi-1.5 is a 1.3 billion parameter Transformer-based language model that demonstrates performance comparable to models 5x larger on natural language tasks. Trained without web data, it surpasses most non-frontier LLMs on complex reasoning tasks such as grade-school mathematics and basic coding. The model exhibits traits like step-by-step thinking but also retains risks including hallucinations and potential for toxic generations.written, reviewed, cited
github.com2026-01-15authorityHN 1 pts · 0 comments2,927 words
Satya is an offline-first educational platform integrating Retrieval-Augmented Generation with the Phi 1.5 language model to serve 1.39 million secondary students in Nepal. The system runs on 4GB RAM hardware using an 800MB model and a local ChromaDB vector database, achieving retrieval in 10-12 seconds with a 512-token context window. It eliminates internet dependency by processing content locally, ensuring access for rural areas where only 17.4% of households are connected.written, reviewed, cited
- The system requires 4GB RAM hardware from 2015 to function without internet dependency.
- Teachers can prepare content for thousands of students with zero ongoing costs after initial setup.
- The diagram service checks a pre-built library before attempting AI generation to ensure grade appropriateness.
emergentmind.com1,044 words
Phi-1.5 is a 1.3B parameter Transformer-based model that leverages textbook-quality synthetic data to enhance reasoning, mathematics, and basic coding. The architecture employs 24 Transformer layers with 32 attention heads and supports a context window of up to 2048 tokens using flash attention. Benchmark evaluations show WinoGrande accuracy of 0.734, matching 7B–13B parameter models, though the model lacks instruction fine-tuning or RLHF.written, reviewed, cited
- Training uses a constant learning rate of 2 × 10^-4 with Adam optimizer momentum parameters 0.9 and 0.98.
- The model is trained entirely in fp16 precision using DeepSpeed ZeRO Stage 2 with overall batch size 2048.
- No instruction fine-tuning or RLHF is applied, meaning out-of-the-box conversation control and safety are limited.
pratyushmaini.github.io1,588 words
debuggercafe.com1,697 words
mlabonne.github.ioabout the phi 1 family
inferencebench.io464 words
Community1
news.ycombinator.com2023-10-05HN 2 pts · 0 comments113 words
