Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -83,6 +83,7 @@ In total there were three different datasets used for alignment:
|
|
| 83 |
**General:**
|
| 84 |
- Token length limits: Embeddings were truncated at 8192 tokens (OpenAI model constraint)
|
| 85 |
- Sentence Length Penalties:
|
|
|
|
| 86 |
**Absolute penalty**:
|
| 87 |
`penalized_sim = cos_sim - α * abs(len_rm - len_de)`
|
| 88 |
|
|
@@ -182,8 +183,11 @@ parallel_data_unfiltered/
|
|
| 182 |
├── matched_{THRESHOLD}_{ALPHA}_{LENGTH_PENALTY_TYPE}.jsonl
|
| 183 |
|
| 184 |
├── matched_0.50_ 0.005_relative.jsonl
|
|
|
|
| 185 |
├── matched_0.50_0.005_absolute.jsonl
|
|
|
|
| 186 |
├── matched_0.50_none.jsonl
|
|
|
|
| 187 |
...
|
| 188 |
|
| 189 |
With ALPHA being `0.005` or `None` and LENGTH_PENALTY_TYPE can be `"relative"`, `"absolute"` or `None`.
|
|
|
|
| 83 |
**General:**
|
| 84 |
- Token length limits: Embeddings were truncated at 8192 tokens (OpenAI model constraint)
|
| 85 |
- Sentence Length Penalties:
|
| 86 |
+
|
| 87 |
**Absolute penalty**:
|
| 88 |
`penalized_sim = cos_sim - α * abs(len_rm - len_de)`
|
| 89 |
|
|
|
|
| 183 |
├── matched_{THRESHOLD}_{ALPHA}_{LENGTH_PENALTY_TYPE}.jsonl
|
| 184 |
|
| 185 |
├── matched_0.50_ 0.005_relative.jsonl
|
| 186 |
+
|
| 187 |
├── matched_0.50_0.005_absolute.jsonl
|
| 188 |
+
|
| 189 |
├── matched_0.50_none.jsonl
|
| 190 |
+
|
| 191 |
...
|
| 192 |
|
| 193 |
With ALPHA being `0.005` or `None` and LENGTH_PENALTY_TYPE can be `"relative"`, `"absolute"` or `None`.
|