harims95/LoopLM-135M-naive

fix: tie_weights accepts kwargs for transformers 1.21+

fix: keep all_tied_weights_keys property pointing to _tied_weights_keys dict

fix: _tied_weights_keys as dict, add tie_weights method

fix forward: always return logits, accept attention_mask, ignore_index=-100 for SFT

expand model card with full architecture details and Parcae journey

fix: all_tied_weights_keys returns dict for transformers 1.21+

fix: add all_tied_weights_keys property for transformers 1.21+

fix: add _tied_weights_keys and output embedding hooks

Initial release: LoopLM-135M-naive trained on FineWeb 4.6B tokens

initial commit