TRM-Text1 (ACT)
TRM-Text1 (ACT) is a causal language model based on a Tiny Recursive Reasoning Model (TRM) with Adaptive Computation Time (ACT) for per-token variable depth.
- Architecture: TRM (causal) + ACT halting
- Training Data: wikitext-103-raw-v1
- Tokenizer: t5-small (SentencePiece)
- Vocab Size: 32100
- Objective: Causal Language Modeling (next-token)
- Seq Len: 1024
Note: This model uses the T5 SentencePiece tokenizer. Perplexity numbers on WT103
reported here are not directly comparable to GPT-2 BPE-based PPLs.
Latest Performance (Epoch 1)
- Validation Loss: 4.8248
- Validation LM Loss: 4.8149
- Validation Ponder Loss: 1.0064
- Validation Perplexity (LM-only): 123.34
