Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
tags:
|
| 4 |
+
- chess
|
| 5 |
+
- sft
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# Chess SFT checkpoints with the positional-encoding fix (20M / 50M / 200M)
|
| 9 |
+
|
| 10 |
+
Multi-turn chess SFT checkpoints trained after the YaRN positional-encoding fix
|
| 11 |
+
in `sft_trainer.py` (2026-09). Before the fix, the `block_size` / `rope_scaling`
|
| 12 |
+
change was applied after `from_pretrained()`, so runs with a 3072-token context
|
| 13 |
+
trained with plain RoPE extrapolation while their saved config claimed YaRN.
|
| 14 |
+
These runs load the config first and build the model from it, so training and
|
| 15 |
+
inference both use YaRN (factor 3, 1024 → 3072). Thanks to Jens Tuyls for spotting the bug.
|
| 16 |
+
|
| 17 |
+
## Setup
|
| 18 |
+
|
| 19 |
+
- SFT data: `200m_generated` (qwen06b generator, nc8/nt10/nta8), CoT format `trajectory_sep_no_labels`, 85 train files
|
| 20 |
+
- Context: 3072 tokens via YaRN on a 1024-token pretrained base
|
| 21 |
+
- LR 3e-4, final checkpoint only (step 771)
|
| 22 |
+
|
| 23 |
+
## Layout
|
| 24 |
+
|
| 25 |
+
Folders mirror the local save tree. Each model lives at `<model_id>/final/`, where
|
| 26 |
+
`model_id = C{compute}_{size}_alpha{α}_beta{β}`. Alongside `final/` are
|
| 27 |
+
`combined_eval.json` (when evaluated) and `config_snapshot/` (training config).
|
| 28 |
+
Optimizer states are not included.
|
| 29 |
+
|
| 30 |
+
```python
|
| 31 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 32 |
+
repo, sub = "Pre-to-Post-2/sft_200m_generated_fixed_pe", "C6p5e18_50m_alpha0.400_beta0.023/final"
|
| 33 |
+
model = AutoModelForCausalLM.from_pretrained(repo, subfolder=sub, trust_remote_code=True)
|
| 34 |
+
tok = AutoTokenizer.from_pretrained(repo, subfolder=sub, trust_remote_code=True)
|
| 35 |
+
```
|
| 36 |
+
|
| 37 |
+
## Models
|
| 38 |
+
|
| 39 |
+
- `C6p5e18_200m_alpha0.100_beta0.100`
|
| 40 |
+
- `C6p5e18_200m_alpha0.200_beta0.100`
|
| 41 |
+
- `C6p5e18_200m_alpha0.750_beta0.100`
|
| 42 |
+
- `C6p5e18_200m_alpha1.000_beta0.100`
|
| 43 |
+
- `C6p5e18_20m_alpha0.010_beta0.008`
|
| 44 |
+
- `C6p5e18_20m_alpha0.030_beta0.008`
|
| 45 |
+
- `C6p5e18_20m_alpha0.050_beta0.008`
|
| 46 |
+
- `C6p5e18_20m_alpha0.100_beta0.008`
|
| 47 |
+
- `C6p5e18_20m_alpha0.200_beta0.008`
|
| 48 |
+
- `C6p5e18_20m_alpha0.300_beta0.008`
|
| 49 |
+
- `C6p5e18_20m_alpha0.600_beta0.008`
|
| 50 |
+
- `C6p5e18_20m_alpha0.750_beta0.008`
|
| 51 |
+
- `C6p5e18_20m_alpha1.000_beta0.008`
|
| 52 |
+
- `C6p5e18_50m_alpha0.010_beta0.023`
|
| 53 |
+
- `C6p5e18_50m_alpha0.030_beta0.023`
|
| 54 |
+
- `C6p5e18_50m_alpha0.050_beta0.023`
|
| 55 |
+
- `C6p5e18_50m_alpha0.100_beta0.023`
|
| 56 |
+
- `C6p5e18_50m_alpha0.200_beta0.023`
|
| 57 |
+
- `C6p5e18_50m_alpha0.400_beta0.023`
|
| 58 |
+
- `C6p5e18_50m_alpha0.750_beta0.023`
|
| 59 |
+
- `C6p5e18_50m_alpha1.000_beta0.023`
|
| 60 |
+
- `C6p5e19_200m_alpha0.200_beta0.007`
|
| 61 |
+
- `C6p5e19_200m_alpha0.400_beta0.007`
|
| 62 |
+
- `C6p5e19_200m_alpha0.750_beta0.007`
|
| 63 |
+
- `C6p5e19_200m_alpha1.000_beta0.007`
|
| 64 |
+
- `C6p5e19_50m_alpha0.180_beta0.002`
|
| 65 |
+
|
| 66 |
+
`C6p5e18_20m_alpha1.000_beta0.008` and `C6p5e18_50m_alpha0.200_beta0.023` were retrained on 2026-10-05 and have no `combined_eval.json` yet.
|