BeardedMonster commited on
Commit
d8e5d37
·
verified ·
1 Parent(s): b50a029

Add judge scores, define columns, explain how scores are derived

Browse files
Files changed (3) hide show
  1. README.md +79 -26
  2. data/test.jsonl.gz +2 -2
  3. data/train.jsonl.gz +2 -2
README.md CHANGED
@@ -20,6 +20,17 @@ dataset_info:
20
  - {name: source_dataset, dtype: string}
21
  - {name: source_register, dtype: string}
22
  - {name: reference, dtype: string}
 
 
 
 
 
 
 
 
 
 
 
23
  - name: candidates
24
  list:
25
  - {name: k, dtype: int64}
@@ -34,6 +45,17 @@ dataset_info:
34
  - {name: system, dtype: string}
35
  - {name: reward, dtype: float64}
36
  - {name: chrf, dtype: float64}
 
 
 
 
 
 
 
 
 
 
 
37
  - name: output
38
  list: string
39
  configs:
@@ -67,7 +89,7 @@ The `subset` column tells the two kinds of rows apart.
67
  **`test` is the benchmark.** It has two parts:
68
 
69
  - **`generated`**: 11,100 English texts, **exactly 300 per language** (37 languages), so no language counts more than another when you average scores. Every text has an `hm` and an `ood` row with eight candidates written by our 1B model (`AletheiaAING/gemma-3-1b-csw-best`). The language pairs and text types follow the same mix as `train` (under 1% difference on source mix, under 0.5% on language pairs), and all 81 base-switch pairs appear, with 66 to 166 `hm` rows each.
70
- - **`curated`**: 6,230 held-out items in 34 languages and five tasks (`eng_hm`, `m_hm`, `eng_ood`, `m_ood`, `hm_ood`, explained under Fields). Each item has a curated `reference` and the outputs of three csw systems with their scores: `ev` (a learned quality score), `reward` (a rule-and-quality reward) and `chrf` (chrF++ against the reference).
71
 
72
  Report scores per language as well as the average.
73
 
@@ -84,28 +106,58 @@ Report scores per language as well as the average.
84
  - Measure progress on African-language code-switching with the `test` split.
85
  - Study how to pick good code-switched text from several candidates.
86
 
87
- ## Fields
88
 
89
- Each line is one (English text, task) pair.
90
 
91
- | Field | Meaning |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92
  |---|---|
93
- | `id` | id of the item; for `generated` rows the `hm` and `ood` rows of one text share it |
94
- | `subset` | `generated` or `curated` (see Train and test) |
95
- | `task` | `hm` (code-switching) or `ood` (idiom-rich rewrite) |
96
- | `task_type` | the full task name: `eng_hm`, `m_hm`, `eng_ood`, `m_ood` or `hm_ood`. `eng_` means the input is English, `m_` that it is already in the base language, and `hm_ood` that it is already code-switched and you want the idiom-rich version. `generated` rows are all `eng_hm` or `eng_ood`. |
97
- | `input_type` | `english`, `base_language` or `code_switched` |
98
- | `lang`, `switch` | base language code; switch language code (`hm` only, otherwise `null`) |
99
- | `pair_in_training_table` | `false` for the 3,348 `hm` rows whose base-switch pair is outside the 67 pairs our 1B model was trained on (see the pair table); `true` otherwise |
100
- | `prompt` | the exact prompt given to the model, for example `<hm> <kon>+<fra>: Hey there! ...` |
101
- | `input` | the text the model receives |
102
- | `english` | the English source text (`null` when the input is not English) |
103
- | `reference` | the curated reference text (`curated` rows only, `null` otherwise) |
104
- | `source_dataset`, `source_register` | where the English text comes from, and its register (`chat`, `story`, `complex`) |
105
- | `candidates` | the candidates, described below |
106
- | `output` | the cleaned text of each candidate, in the same order as `candidates`: language tags and the `<dev>...</dev>` block removed, so you can use it directly |
107
-
108
- Each candidate has `k` (0 is the greedy output, 1 and up are samples), `greedy`, `text` (the raw output with tags and device block), `finished` (stopped by itself, not cut off), `devices` (the switched phrases or idioms), `final` (the text without tags and device block), `format_ok` (one device block and a non-empty finished text) and `ev` (a learned quality score, `null` when missing), and for `curated` rows also `system` (`sft`, `rl_best` or `rl_last`), `reward` and `chrf`. Fields that don't apply are `null`. Use `output[i]` for the text of `candidates[i]`.
 
 
 
 
 
 
 
109
 
110
  Example row, shortened:
111
 
@@ -321,13 +373,14 @@ Rows are (English text, task) pairs; every generated text has an `hm` and an `oo
321
  | Zulu | `zul` | 9 | 9 | 9 | 9 | 9 | 45 |
322
  | **Total** | | 1190 | 1190 | 1359 | 1359 | 1132 | 6230 |
323
 
324
- **Scores of the three systems on the curated items** (mean over 6,230 items; `ev` and `reward` come from learned models and rules, not from people).
325
 
326
- | `system` | what it is | items | `ev` | `chrf` |
327
- |---|---|---|---|---|
328
- | `sft` | supervised | 6,225 | 3.53 | 43.9 |
329
- | `rl_best` | reinforcement-learned, best checkpoint | 6,225 | 3.75 | 43.5 |
330
- | `rl_last` | reinforcement-learned, last checkpoint | 6,225 | 3.73 | 43.5 |
 
331
 
332
  ## Supported prompt formats
333
 
 
20
  - {name: source_dataset, dtype: string}
21
  - {name: source_register, dtype: string}
22
  - {name: reference, dtype: string}
23
+ - name: reference_judge
24
+ struct:
25
+ - {name: meaning, dtype: float64}
26
+ - {name: base_language, dtype: float64}
27
+ - {name: switch_language, dtype: float64}
28
+ - {name: language_compliance, dtype: float64}
29
+ - {name: switching_naturalness, dtype: float64}
30
+ - {name: language_correctness, dtype: float64}
31
+ - {name: idiom_use, dtype: float64}
32
+ - {name: naturalness, dtype: float64}
33
+ - {name: overall, dtype: float64}
34
  - name: candidates
35
  list:
36
  - {name: k, dtype: int64}
 
45
  - {name: system, dtype: string}
46
  - {name: reward, dtype: float64}
47
  - {name: chrf, dtype: float64}
48
+ - name: judge
49
+ struct:
50
+ - {name: meaning, dtype: float64}
51
+ - {name: base_language, dtype: float64}
52
+ - {name: switch_language, dtype: float64}
53
+ - {name: language_compliance, dtype: float64}
54
+ - {name: switching_naturalness, dtype: float64}
55
+ - {name: language_correctness, dtype: float64}
56
+ - {name: idiom_use, dtype: float64}
57
+ - {name: naturalness, dtype: float64}
58
+ - {name: overall, dtype: float64}
59
  - name: output
60
  list: string
61
  configs:
 
89
  **`test` is the benchmark.** It has two parts:
90
 
91
  - **`generated`**: 11,100 English texts, **exactly 300 per language** (37 languages), so no language counts more than another when you average scores. Every text has an `hm` and an `ood` row with eight candidates written by our 1B model (`AletheiaAING/gemma-3-1b-csw-best`). The language pairs and text types follow the same mix as `train` (under 1% difference on source mix, under 0.5% on language pairs), and all 81 base-switch pairs appear, with 66 to 166 `hm` rows each.
92
+ - **`curated`**: 6,230 held-out items in 34 languages and five tasks (`eng_hm`, `m_hm`, `eng_ood`, `m_ood`, `hm_ood`, explained under Columns). Each item has a curated `reference` and the outputs of three csw systems, each with four scores: `ev`, `reward`, `chrf` and `judge` (how they are derived is explained under How the scores were derived). The reference has judge grades too.
93
 
94
  Report scores per language as well as the average.
95
 
 
106
  - Measure progress on African-language code-switching with the `test` split.
107
  - Study how to pick good code-switched text from several candidates.
108
 
109
+ ## Columns
110
 
111
+ Each row is one input text with one task. Columns that don't apply to a row are `null`.
112
 
113
+ **Row columns**
114
+
115
+ | Column | Meaning |
116
+ |---|---|
117
+ | `id` | id of the item. For `generated` rows the `hm` and `ood` rows of one text share it. |
118
+ | `subset` | `generated` (English texts with candidates written by our 1B model) or `curated` (the held-out test items with a reference, see Train and test). |
119
+ | `task` | what is asked: `hm` (code-switching) or `ood` (idiom-rich rewrite). |
120
+ | `task_type` | the full task name: `eng_hm`, `m_hm`, `eng_ood`, `m_ood` or `hm_ood`. `eng_` means the input is English, `m_` means it is already in the base language, and `hm_ood` means it is already code-switched and you want the idiom-rich version. `generated` rows are all `eng_hm` or `eng_ood`. |
121
+ | `input_type` | the language of the input: `english`, `base_language` or `code_switched`. |
122
+ | `lang` | base language code, for example `yor`. |
123
+ | `switch` | switch language code for `hm` tasks and `hm_ood`, otherwise `null`. |
124
+ | `pair_in_training_table` | `false` for the 3,348 `hm` rows whose base-switch pair is outside the 67 pairs our 1B model was trained on (see the pair table), `true` otherwise. Use it to keep only the supported pairs. |
125
+ | `prompt` | the exact prompt given to the model, for example `<hm> <kon>+<fra>: Hey there! ...`. |
126
+ | `input` | the text the model had to rewrite (also inside `prompt`). |
127
+ | `english` | the English source text, `null` when the input is not English. |
128
+ | `source_dataset`, `source_register` | where the English text comes from and its register (`chat`, `story`, `complex`); `curated` rows have `source_dataset = curated`. |
129
+ | `reference` | the curated reference rewrite, with tags and device block removed. `curated` rows only. |
130
+ | `reference_judge` | judge grades of the `reference` (see below). `curated` rows only. |
131
+ | `candidates` | the candidate outputs, one entry each (columns below). |
132
+ | `output` | the cleaned text of each candidate, in the same order as `candidates`: language tags and the `<dev>...</dev>` block removed, so you can use it directly. |
133
+
134
+ **Columns inside each entry of `candidates`**
135
+
136
+ | Column | Meaning |
137
  |---|---|
138
+ | `k` | candidate number. For `generated` rows 0 is the greedy output and 1 to 7 are samples; for `curated` rows it numbers the three systems. |
139
+ | `greedy` | `true` if this is a greedy (temperature 0) output. |
140
+ | `text` | the raw output: language tags, `<dev>` block and text. |
141
+ | `finished` | `true` if the model stopped by itself and was not cut off at the length limit (`null` for `curated` rows). |
142
+ | `devices` | the switched phrases (`hm`) or idioms (`ood`) the model listed, as they appear in the text. |
143
+ | `final` | the text without tags and device block. |
144
+ | `format_ok` | `true` if the output has the language tags, one device block and a non-empty text. |
145
+ | `ev` | learned quality score, 1 to 10, `null` when missing. |
146
+ | `system` | `curated` rows only: `sft` (supervised), `rl_best` or `rl_last` (reinforcement-learned, best and last checkpoint). |
147
+ | `reward` | `curated` rows only: the rule-based reward (can be negative; higher is better). |
148
+ | `chrf` | `curated` rows only: chrF++ of `final` against `reference`, 0 to 100. |
149
+ | `judge` | `curated` rows only: judge grades of this output, a record with the criteria below. |
150
+
151
+ Use `output[i]` for the text of `candidates[i]`.
152
+
153
+ ## How the scores were derived
154
+
155
+ None of these scores comes from a person.
156
+
157
+ - **`ev`**: a learned quality model (a multilingual encoder we fine-tuned on rated code-switched outputs). It reads the prompt and an output and predicts a 1 to 10 quality score. It is only computed for outputs with the right format. It likes short, simple code-switching, so don't use it as the only filter.
158
+ - **`reward`**: the reward used to train the two reinforcement-learned systems. It is the learned quality score combined with a format check and rule penalties for things like unused devices, the wrong language, repetition, wrong length, English leaking in and copying the input.
159
+ - **`chrf`**: sentence-level chrF++ (word order 2, sacreBLEU) between the cleaned output and the cleaned `reference`. It measures closeness to the reference, not quality on its own.
160
+ - **`judge`, `reference_judge`**: grades from 1 to 5 (mostly whole numbers) given by a large-language-model judge with a fixed rubric. The judge saw the input and the output but not which system wrote it. Anchors: 5 means a native speaker would change nothing, 1 means unusable. `hm` rows are graded on `meaning`, `base_language`, `switch_language`, `language_compliance`, `switching_naturalness` and `overall`; `ood` rows on `meaning`, `language_correctness`, `idiom_use`, `naturalness`, `language_compliance` and `overall`. A criterion the judge could not assess is `null`, and criteria that don't apply to the task are `null`. A few outputs could not be parsed and have no grades.
161
 
162
  Example row, shortened:
163
 
 
373
  | Zulu | `zul` | 9 | 9 | 9 | 9 | 9 | 45 |
374
  | **Total** | | 1190 | 1190 | 1359 | 1359 | 1132 | 6230 |
375
 
376
+ **Scores of the three systems on the curated items** (mean over the curated items; none of these scores comes from a person; see How the scores were derived).
377
 
378
+ | `system` | what it is | items | `ev` (1-10) | `chrf` (0-100) | judge `overall` (1-5) |
379
+ |---|---|---|---|---|---|
380
+ | `sft` | supervised | 6,225 | 3.53 | 43.9 | 2.27 |
381
+ | `rl_best` | reinforcement-learned, best checkpoint | 6,225 | 3.75 | 43.5 | 2.15 |
382
+ | `rl_last` | reinforcement-learned, last checkpoint | 6,225 | 3.73 | 43.5 | 2.15 |
383
+ | `reference` | the curated reference text | 6,225 | n/a | n/a | 3.99 |
384
 
385
  ## Supported prompt formats
386
 
data/test.jsonl.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:50451bb639423a5605a65171a074f41aecc87b317f84c2a0dcacfb68841d17ce
3
- size 42844020
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ad925e90b0cad47deb9700fee72992c3b310811289485255eb6f0cba3fa240c
3
+ size 43110091
data/train.jsonl.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5b02433866e008fb96bd9136359e4f1e6c52fcbf847d9a3024d67363d14b8b20
3
- size 29295815
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7245a0b644c785f53a44db649cb257754ed3ffe8d4e5893678267e9c5cb66155
3
+ size 29324473