Add model card: v1 historical TinyLM checkpoint
Browse files
README.md
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
tags:
|
| 5 |
+
- causal-lm
|
| 6 |
+
- mla
|
| 7 |
+
- muon
|
| 8 |
+
- historical
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# TinyLM v1 Checkpoint (historical)
|
| 12 |
+
|
| 13 |
+
Single MLA+Muon training run from the v1 TinyLM effort (RunPod A100-80GB,
|
| 14 |
+
May 2026). Trained on **1B unique** FineWeb-Edu tokens **repeated ~21×**
|
| 15 |
+
over 20k steps — the data bug the HPC re-run later fixed.
|
| 16 |
+
|
| 17 |
+
**Preserved here for historical contrast — not the recommended model.**
|
| 18 |
+
|
| 19 |
+
- **Recommended model:** [`Shiv-22/tinylm`](https://huggingface.co/Shiv-22/tinylm) (Run D from the HPC re-run, 8B unique tokens, +3.97 avg pts above this v1 on the same architecture)
|
| 20 |
+
- **Full 4-arm ablation checkpoints:** [`Shiv-22/tinylm-checkpoints-v2`](https://huggingface.co/Shiv-22/tinylm-checkpoints-v2)
|
| 21 |
+
- **Source code:** [github.com/shivnarainms22/TinyLM](https://github.com/shivnarainms22/TinyLM)
|
| 22 |
+
|
| 23 |
+
## v1 eval (0-shot)
|
| 24 |
+
|
| 25 |
+
| Benchmark | Metric | v1 D |
|
| 26 |
+
|---|---|:---:|
|
| 27 |
+
| HellaSwag | acc_norm | 37.1% |
|
| 28 |
+
| ARC-Easy | acc_norm | 48.4% |
|
| 29 |
+
| LAMBADA | acc | 29.2% |
|
| 30 |
+
| Winogrande | acc | 50.0% |
|
| 31 |
+
| **Average** | | **41.18%** |
|
| 32 |
+
|
| 33 |
+
The notably weak LAMBADA (long-range coherence) was the main signal that
|
| 34 |
+
repeated data was hurting; the HPC re-run with 8B unique tokens lifted
|
| 35 |
+
LAMBADA to 36.81% (+7.61) on the same arm.
|
| 36 |
+
|
| 37 |
+
## License
|
| 38 |
+
|
| 39 |
+
Apache 2.0.
|