lilm1-200M

Model Details

• Architecture: Hybrid LFM (10 Convolutional + 6 Attention layers)
• Parameters: 262.2M
• Vocab Size: 49152 (using tokenizer HuggingFaceTB/SmolLM2-135M)
• Sequence Length: 2048
• Dataset: glouriousgautam/lilm-pretrain-12B (FineWeb, FineWeb-Edu, FineMath, DCLM, FinePDFs)
• Total Tokens: ~10.5B tokens
• Precision: bfloat16

Training Configuration

• Effective Batch Size: 294,912 tokens/step
• Total Training Steps: 32000
• Warmup Steps: 640
• Muon Optimizer: LR = 0.01, Momentum = 0.95, WD = 0.01
• AdamW Optimizer: LR = 0.0003, WD = 0.1, Betas = (0.9, 0.95)

Metrics & Tracking

• Final Training Loss: 2.7235
• Final Perplexity: 15.23
• Tokens Seen: 9,692,774,400
• Weights & Biases Project: lilm-pretraining