SmolLM2-135M
PulseAugur coverage of SmolLM2-135M — every cluster mentioning SmolLM2-135M across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
-
Small language model SmolLM2-135M shows strong understanding despite hallucination issues
A user is testing the SmolLM2-135M model on their phone and observed an interesting failure mode. The model correctly understands the question and begins with the right answer, but then hallucinates the remainder of its…
-
Research paper proposes synthetic data verification to prevent model collapse
A new research paper explores the phenomenon of "model collapse," where generative models trained on their own synthetic data degrade in performance over time. The study proposes that incorporating an external synthetic…
-
Small language model trained on single GPU detailed in new study
Researchers have detailed a method for training a small language model, L20-Edu-135M, using significantly fewer computational resources, specifically on a single NVIDIA L20 GPU. The study focused on data efficiency, uti…
-
LLM Tutorial Explains Foundation, Instruct, and Chat Model Differences
A tutorial demonstrates the distinctions between foundation, instruct, and chat models in large language models. It uses the SmolLM2-135M family, runnable on Google Colab without a GPU, to illustrate how models evolve f…
-
Compact Bangla LLM Outperforms Larger Models with Efficient Design
Researchers have developed a new compact language model, bangla-smollm-135m, specifically designed for the Bangla language. This 135-million parameter model achieves competitive performance against larger models by empl…
-
AutoMegaKernel compiles Llama models into single CUDA kernels
Researchers have developed AutoMegaKernel (AMK), a system that compiles HuggingFace Llama-family models into a single, persistent CUDA kernel for efficient forward passes. AMK's static validator ensures schedule safety,…
-
Jetson Orin Nano benchmarks 8 tiny LLMs across power modes
A benchmark of eight small language models (135M to ~1B parameters) was conducted on a Jetson Orin Nano Super 8GB device. The tests explored four power modes (7W, 15W, 25W, MAXN) using the llama.cpp CUDA backend. The fi…
-
FlashNorm speeds up transformer inference by optimizing normalization layers
Researchers have developed FlashNorm, a technique to accelerate normalization layers in Transformer models. By reformulating RMSNorm and folding its weights into subsequent linear layers, FlashNorm enables parallel exec…