Datasets:
docs: name the real config for conditions 3 and 4 (phase-2 namespace), fix the load example
Browse filesThe card's config table, 'Used by' lines and usage example referred to condition-3-zh-5k and condition-4-zh-5k, which are not configs on this repo: the only cond-3/cond-4 configs are phase-2-the-stack-v1-condition-{3,4}-zh-5k. Name those, and extend the 2026-05-18 note to say the prefix is a filing location, not a Stack v1 provenance claim. Context: the cond-3 seed-123 adapter was trained from that config and inherited the phase-2 name on upload (language-decoded-lora PR #17 renames it). No data files change.
README.md
CHANGED
|
@@ -917,7 +917,7 @@ dataset_info:
|
|
| 917 |
|
| 918 |
> **Experiment and proposed paper title:** _Language Decoded: Exploring the Impact of Native Code on Multilingual Models_
|
| 919 |
|
| 920 |
-
> **Note (2026-05-18):** Current Phase 3 configs use the short `condition-*` namespace and include `103k`, `20k`, and `5k` sizes for Conditions 1--2. Phase 2 configs remain available under the `phase-2-the-stack-v1-*` namespace for reproducibility.
|
| 921 |
|
| 922 |
Multilingual Python code datasets for the **Language Decoded** project (part of [Cohere's Tiny Aya Expedition](https://aya.for.ai)). See [legesher/language-decoded-experiments](https://huggingface.co/datasets/legesher/language-decoded-experiments) for the canonical project description, the full experimental ladder, and the paper-grade evaluation results.
|
| 923 |
|
|
@@ -964,8 +964,8 @@ Conditions 1--2 are available in three current Phase 3 sizes: `-103k` full corpo
|
|
| 964 |
| `condition-2-ur-103k` | 2 | Urdu | Legesher v0.7.3 reserved-word translation | 93,547 | 10,395 |
|
| 965 |
| `condition-2-ur-20k` | 2 | Urdu | Random 20k subset of `condition-2-ur-103k` | 18,000 | 2,000 |
|
| 966 |
| `condition-2-ur-5k` | 2 | Urdu | Compact 5k subset | 4,500 | 500 |
|
| 967 |
-
| `condition-3-zh-5k`
|
| 968 |
-
| `condition-4-zh-5k`
|
| 969 |
| `condition-5-ur-5k-c4ai-aya-expanse-32b` | 5 | Urdu | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,088 | 381 |
|
| 970 |
| `condition-5-zh-5k-c4ai-aya-expanse-32b` | 5 | Chinese | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,052 | 381 |
|
| 971 |
| `condition-5-es-5k-c4ai-aya-expanse-32b` | 5 | Spanish | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,032 | 329 |
|
|
@@ -1003,7 +1003,7 @@ Condition 5 uses the conditions 1--2 schema plus an `idx` column. `code` is the
|
|
| 1003 |
|
| 1004 |
### Condition 3
|
| 1005 |
|
| 1006 |
-
Used by: `condition-3-zh-5k`
|
| 1007 |
|
| 1008 |
In Phase 3, Condition 3 ("Mixed Native Sources") refers to **community-collected raw Chinese code from varied online public-source repositories** — reflecting how non-English Python is actually used in real-world projects. The "Mixed Native Sources" name carries from Phase 2, where it originally referred to a planned composite (native code padded with cond-2 transpiled files); in Phase 3 the "mixed" refers to the diversity of source locations, not a cond-2/native composite. The physical dataset has not changed across phases.
|
| 1009 |
|
|
@@ -1021,7 +1021,7 @@ The schema includes a `source_type` column from the Phase 2 composite design, wh
|
|
| 1021 |
|
| 1022 |
### Condition 4
|
| 1023 |
|
| 1024 |
-
Used by: `condition-4-zh-5k`
|
| 1025 |
|
| 1026 |
Condition 4 ("Community-Contributed Native Code") is intended to contain code whose problem-solving logic is itself native — written as if a native speaker were approaching the problem, not English code that was later translated. The current dataset reflects an earlier Phase 2 attempt to assemble this corpus; community contributions were insufficient for stable training, so cond-4 was not evaluated in either Phase 2 or Phase 3. Cond-5's fully-translated data served as Phase 3's practical proxy because gathering native-authored code at scale proved difficult. Direct contributions to the cond-4 corpus are open at the [`legesher/legesher-native-code`](https://huggingface.co/spaces/legesher/legesher-native-code) HF Space.
|
| 1027 |
|
|
@@ -1084,10 +1084,11 @@ ds = load_dataset("legesher/language-decoded-data", "condition-2-es-5k")
|
|
| 1084 |
ds = load_dataset("legesher/language-decoded-data", "condition-2-ur-5k")
|
| 1085 |
|
| 1086 |
# Load blended native + transpiled (condition 3)
|
| 1087 |
-
|
|
|
|
| 1088 |
|
| 1089 |
# Load strictly native code (condition 4)
|
| 1090 |
-
ds = load_dataset("legesher/language-decoded-data", "condition-4-zh-5k")
|
| 1091 |
|
| 1092 |
# Load model-translated code (condition 5 -- raw, pre-cleanup)
|
| 1093 |
ds = load_dataset("legesher/language-decoded-data", "condition-5-ur-5k-c4ai-aya-expanse-32b")
|
|
|
|
| 917 |
|
| 918 |
> **Experiment and proposed paper title:** _Language Decoded: Exploring the Impact of Native Code on Multilingual Models_
|
| 919 |
|
| 920 |
+
> **Note (2026-05-18, updated 2026-10-13):** Current Phase 3 configs use the short `condition-*` namespace and include `103k`, `20k`, and `5k` sizes for Conditions 1--2, plus the three `condition-5-*` configs. Phase 2 configs remain available under the `phase-2-the-stack-v1-*` namespace for reproducibility. **Conditions 3 and 4 have no short-name config:** their corpora are community-collected native code, unchanged across phases, and are loaded as `phase-2-the-stack-v1-condition-3-zh-5k` and `phase-2-the-stack-v1-condition-4-zh-5k`. The namespace prefix reflects where the files are filed, not their source (they do not come from The Stack v1). The Phase 3 cond-3 adapters were trained from that config.
|
| 921 |
|
| 922 |
Multilingual Python code datasets for the **Language Decoded** project (part of [Cohere's Tiny Aya Expedition](https://aya.for.ai)). See [legesher/language-decoded-experiments](https://huggingface.co/datasets/legesher/language-decoded-experiments) for the canonical project description, the full experimental ladder, and the paper-grade evaluation results.
|
| 923 |
|
|
|
|
| 964 |
| `condition-2-ur-103k` | 2 | Urdu | Legesher v0.7.3 reserved-word translation | 93,547 | 10,395 |
|
| 965 |
| `condition-2-ur-20k` | 2 | Urdu | Random 20k subset of `condition-2-ur-103k` | 18,000 | 2,000 |
|
| 966 |
| `condition-2-ur-5k` | 2 | Urdu | Compact 5k subset | 4,500 | 500 |
|
| 967 |
+
| `phase-2-the-stack-v1-condition-3-zh-5k` (no short name; see note above) | 3 | Chinese | Blended: native Chinese code + transpiled Python | 4,500 | 500 |
|
| 968 |
+
| `phase-2-the-stack-v1-condition-4-zh-5k` (no short name; see note above) | 4 | Chinese | Strictly native Chinese code | 6,553 | 729 |
|
| 969 |
| `condition-5-ur-5k-c4ai-aya-expanse-32b` | 5 | Urdu | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,088 | 381 |
|
| 970 |
| `condition-5-zh-5k-c4ai-aya-expanse-32b` | 5 | Chinese | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,052 | 381 |
|
| 971 |
| `condition-5-es-5k-c4ai-aya-expanse-32b` | 5 | Spanish | Model-translated (full LLM translation via Cohere Aya) — raw, pre-cleanup | 4,032 | 329 |
|
|
|
|
| 1003 |
|
| 1004 |
### Condition 3
|
| 1005 |
|
| 1006 |
+
Used by: `phase-2-the-stack-v1-condition-3-zh-5k` (the only cond-3 config; there is no short-name alias)
|
| 1007 |
|
| 1008 |
In Phase 3, Condition 3 ("Mixed Native Sources") refers to **community-collected raw Chinese code from varied online public-source repositories** — reflecting how non-English Python is actually used in real-world projects. The "Mixed Native Sources" name carries from Phase 2, where it originally referred to a planned composite (native code padded with cond-2 transpiled files); in Phase 3 the "mixed" refers to the diversity of source locations, not a cond-2/native composite. The physical dataset has not changed across phases.
|
| 1009 |
|
|
|
|
| 1021 |
|
| 1022 |
### Condition 4
|
| 1023 |
|
| 1024 |
+
Used by: `phase-2-the-stack-v1-condition-4-zh-5k` (the only cond-4 config; there is no short-name alias)
|
| 1025 |
|
| 1026 |
Condition 4 ("Community-Contributed Native Code") is intended to contain code whose problem-solving logic is itself native — written as if a native speaker were approaching the problem, not English code that was later translated. The current dataset reflects an earlier Phase 2 attempt to assemble this corpus; community contributions were insufficient for stable training, so cond-4 was not evaluated in either Phase 2 or Phase 3. Cond-5's fully-translated data served as Phase 3's practical proxy because gathering native-authored code at scale proved difficult. Direct contributions to the cond-4 corpus are open at the [`legesher/legesher-native-code`](https://huggingface.co/spaces/legesher/legesher-native-code) HF Space.
|
| 1027 |
|
|
|
|
| 1084 |
ds = load_dataset("legesher/language-decoded-data", "condition-2-ur-5k")
|
| 1085 |
|
| 1086 |
# Load blended native + transpiled (condition 3)
|
| 1087 |
+
# Conditions 3 and 4 live under the phase-2 namespace (corpus unchanged across phases)
|
| 1088 |
+
ds = load_dataset("legesher/language-decoded-data", "phase-2-the-stack-v1-condition-3-zh-5k")
|
| 1089 |
|
| 1090 |
# Load strictly native code (condition 4)
|
| 1091 |
+
ds = load_dataset("legesher/language-decoded-data", "phase-2-the-stack-v1-condition-4-zh-5k")
|
| 1092 |
|
| 1093 |
# Load model-translated code (condition 5 -- raw, pre-cleanup)
|
| 1094 |
ds = load_dataset("legesher/language-decoded-data", "condition-5-ur-5k-c4ai-aya-expanse-32b")
|