Fix v4 token-count arithmetic: v4 data build is 257.9M tok (fw 150.6M + dclm 107.3M), cumulative ~308M, ~30 tok/param, 8.4x shortfall — not the previously stated ~393M/~444M/~43.1. Verified against models/ldt-10m-v4/train.log data-build lines.

Update to v4 checkpoint: 444M tokens, val 3.8943 (ppl 49.12), still incoherent (#11)

v4 checkpoint: 444M tokens, val 3.8943 (ppl 49.12)

Fix fabricated sample quote + overstatement: replace sample #1 (not in eval output) with real samples from eval_ldt10m_final.json; correct "coherent/fluency" to "grammatically structured but semantically incoherent" (#10)

Add LDT-10M config (#9)

Add LDT-10M weights (10,284,480 params, 47 tensors, tied embeddings) (#8)

Add LDT-10M model card (#7)

Add honest model card (architecture, training, undertraining status, real samples)

Tokenizer config (#5)

Architecture config (#4)

Byte-level BPE tokenizer (vocab 12288) (#3)

Self-contained LDT model definition (#2)

LDT-10M weights (47 tensors, 10,284,480 params, tied embeddings) (#1)

initial commit