kaptaan45/QaptaanLM-0.75B

docs: update README with latest architectural specs, quantization tables, and generation configs

Update generation_config.json with optimal greedy decoding, repetition_penalty=1.10, and FIM EOS tokens

Update generation_config.json: Optimized greedy decoding with repetition_penalty=1.12 and multi-EOS token IDs for stable code completion

Fix EOS/PAD token IDs to 248044 (<|endoftext|>)

Publish ultra-detailed comprehensive model card with full architecture and specifications

Update Model Card with detailed technical specifications and professional formatting

Set verified optimal CPT code generation config (temp=0.45, top_p=0.90, rep_pen=1.10)

Set optimal base model code generation config (greedy, rep_pen=1.05)

Update model.safetensors: enforce tied embeddings (752M parameters)

Update README.md

Update comprehensive model card for QaptaanLM-0.75B

Fix QaptaanCache standalone class & mask handling

Enable fast O(1) single-token recurrent caching (QaptaanCache)

Sync official Qwen3.5 merges.txt (12.8MB tokenizer)

Sync official Qwen3.5 vocab.json (12.8MB tokenizer)

Sync official Qwen3.5 tokenizer_config.json (12.8MB tokenizer)

Sync official Qwen3.5 tokenizer.json (12.8MB tokenizer)

Upload model weights and configuration files

Initial release: QaptaanLM-0.75B-Base (1B CPT on KapCode-1B)

initial commit