LLaMA-1B-dj-refine-150B

Use cases

• Edge inference on CPU or low-VRAM devices
• Benchmarking data-cleaning pipelines for LLM pretraining
• Studying effect of data quality on small-model perplexity
• Prototype text-generation features before scaling up

Pros

• Apache-2.0 license — no usage restrictions
• Tiny footprint fits in under 2 GB RAM at fp16
• Training data lineage is documented via Data-Juicer repo
• Compatible with standard Transformers text-generation pipeline

Cons

• 1B parameters produces noticeably weaker reasoning than 7B+ models
• Instruction-following is absent — requires fine-tuning for chat use
• Outperformed on most benchmarks by Qwen2-0.5B despite smaller size
• No GGUF or quantized variants from the original author

Tags