Search

Few direct matches — filled in with the latest updates.

Tag: #local-hardware3 results

🔬

GRPO Trains Qwen for Reddit Summaries on Mac Minis

Trained Qwen2.5-0.5B-Instruct variants on Reddit summarization with length constraints using GRPO on 3x Mac Minis. Quality + length penalty outperformed length-only in LLM-as-a-Judge evals (2.5/4 vs 2.4/4) via DeepEval on faithfulness, coverage, conciseness, clarity. Significant p=0.0042 on ROUGE-L composite score from 200 smoltldr test samples.

Reddit r/LocalLLaMACommunityApr 16#fine-tuning#summarization#local-hardware
LLMs Transformed in One Year

LLMs Transformed in One Year

Post highlights rapid progress in models like Kimi, Minimax, Qwen, Gemma, GLM enabling local runs on affordable hardware. Benchmarks show trends beyond numbers; anticipates Qwen 3.6 27B release. Users now complete complex tasks locally rivaling Claude or GPT-4o.

Reddit r/LocalLLaMACommunityApr 21#model-progress#local-hardware#open-models