Remove unsolicited benchmark commentary
Browse files
README.md
CHANGED
|
@@ -57,8 +57,6 @@ russian = load_dataset("apsua/MIRON_Benchmark", "ru", split="train")
|
|
| 57 |
english = load_dataset("apsua/MIRON_Benchmark", "en", split="train")
|
| 58 |
```
|
| 59 |
|
| 60 |
-
The `train` label is the file split exposed by the dataset loader; the dataset is intended as a benchmark, so avoid training on examples used for evaluation. Select a language configuration when loading the dataset.
|
| 61 |
-
|
| 62 |
## Evaluation
|
| 63 |
|
| 64 |
The benchmark was designed around two complementary measures:
|
|
@@ -66,12 +64,6 @@ The benchmark was designed around two complementary measures:
|
|
| 66 |
- **Generation score:** character-level Levenshtein similarity between a generated continuation and the reference, expressed as a percentage. For two empty strings, the similarity is 100%.
|
| 67 |
- **Target confidence:** geometric mean of the model probabilities assigned to the reference continuation tokens, conditioned on the prefix and preceding reference tokens, expressed as a percentage.
|
| 68 |
|
| 69 |
-
For reproducible comparisons, report results separately by language and category, and specify the model, tokenizer, generation settings and text normalization. Token boundaries can change when `prefix` and `target` are joined; use a consistent, documented procedure for identifying target tokens. Confidence is tokenizer-dependent and should not be interpreted as a calibrated probability that an answer is correct.
|
| 70 |
-
|
| 71 |
-
## Limitations
|
| 72 |
-
|
| 73 |
-
This benchmark covers a small, selected set of short continuation tasks. It does not measure general intelligence or establish the cause of an error: low scores can reflect the data, tokenizer, model or evaluation procedure. Scores are most useful when paired with task-level analysis and other evaluations.
|
| 74 |
-
|
| 75 |
## License
|
| 76 |
|
| 77 |
MIT, as specified in the dataset metadata.
|
|
|
|
| 57 |
english = load_dataset("apsua/MIRON_Benchmark", "en", split="train")
|
| 58 |
```
|
| 59 |
|
|
|
|
|
|
|
| 60 |
## Evaluation
|
| 61 |
|
| 62 |
The benchmark was designed around two complementary measures:
|
|
|
|
| 64 |
- **Generation score:** character-level Levenshtein similarity between a generated continuation and the reference, expressed as a percentage. For two empty strings, the similarity is 100%.
|
| 65 |
- **Target confidence:** geometric mean of the model probabilities assigned to the reference continuation tokens, conditioned on the prefix and preceding reference tokens, expressed as a percentage.
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
## License
|
| 68 |
|
| 69 |
MIT, as specified in the dataset metadata.
|