Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model: allenai/DataDecide-dolma1_7-no-flan-300M
|
| 6 |
+
tags:
|
| 7 |
+
- math
|
| 8 |
+
- worked-examples
|
| 9 |
+
- continual-pretraining
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Same Content, Different Format: worked-example models
|
| 13 |
+
|
| 14 |
+
The trained checkpoints of the worked-example experiment of *Same Content, Different Format: QA Augmentation and
|
| 15 |
+
Faded Worked Examples Under a Fixed Training Budget*. Each subfolder is one arm, continually pretrained from the
|
| 16 |
+
public DataDecide dolma1_7 no-FLAN 300M checkpoint (an OLMo 1 architecture, 371M parameters), converted to the
|
| 17 |
+
Hugging Face Llama format.
|
| 18 |
+
|
| 19 |
+
| Subfolder | Arm | Passes over its corpus |
|
| 20 |
+
|---|---|---|
|
| 21 |
+
| `bare` | problem and final answer only | 298 |
|
| 22 |
+
| `complete` | problem, full step-by-step solution and final answer | 97 |
|
| 23 |
+
| `fade_ordered` | faded solutions, support removed in decreasing order | 19 |
|
| 24 |
+
| `fade_shuffled` | the same faded documents in random order | 19 |
|
| 25 |
+
| `bare_pm` | bare, trained for the fade arms' pass count | 19 |
|
| 26 |
+
| `complete_pm` | complete, trained for the fade arms' pass count | 19 |
|
| 27 |
+
|
| 28 |
+
```python
|
| 29 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 30 |
+
tok = AutoTokenizer.from_pretrained("nzhao721/p1-worked-examples", subfolder="complete")
|
| 31 |
+
model = AutoModelForCausalLM.from_pretrained("nzhao721/p1-worked-examples", subfolder="complete")
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
The training data and the evaluation holdout are in
|
| 35 |
+
[nzhao721/p1-worked-examples-data](https://huggingface.co/datasets/nzhao721/p1-worked-examples-data). Every arm was
|
| 36 |
+
trained once with one seed, so results should be read as directional.
|