Model card updated after epoch 0
Browse files
README.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: t5-small
|
| 3 |
+
tags: [trm, act, recursive, text-generation, wikitext]
|
| 4 |
+
metrics: [loss, lm_loss, ponder_loss, perplexity_lm]
|
| 5 |
+
---
|
| 6 |
+
# TRM-Text1 (ACT)
|
| 7 |
+
|
| 8 |
+
**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.
|
| 9 |
+
|
| 10 |
+
- **Architecture:** TRM (causal) + ACT halting
|
| 11 |
+
- **Training Data:** wikitext-103-raw-v1
|
| 12 |
+
- **Tokenizer:** t5-small (SentencePiece)
|
| 13 |
+
- **Vocab Size:** 32100
|
| 14 |
+
- **Objective:** Causal Language Modeling (next-token)
|
| 15 |
+
- **Seq Len:** 1024
|
| 16 |
+
|
| 17 |
+
Note: This model uses the T5 SentencePiece tokenizer. Perplexity numbers on WT103
|
| 18 |
+
reported here are not directly comparable to GPT-2 BPE-based PPLs.
|
| 19 |
+
|
| 20 |
+
### Latest Performance (Epoch 0)
|
| 21 |
+
- **Validation Loss**: 4.8829
|
| 22 |
+
- **Validation LM Loss**: 4.8728
|
| 23 |
+
- **Validation Ponder Loss**: 1.0091
|
| 24 |
+
- **Validation Perplexity (LM-only)**: 130.69
|