Upload folder using huggingface_hub

README.md

@@ -17,7 +17,7 @@ The original NVIDIA dataset blends multiple sources. For this parallel corpus, w
 | **[Skywork/Skywork-OR1-RL-Data](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data)** (0.12 ratio) | `nano_v3_sft_profiled_skywork_no_omni.jsonl` | 2,998 | 2,975 | 2,979 |
 | **[nvidia/Nemotron-RL-knowledge-mcqa](https://huggingface.co/datasets/nvidia/Nemotron-RL-knowledge-mcqa)** (0.20 ratio) | `nano_v3_sft_profiled_stem_mcqa.jsonl` | 2,639 | 2,850 | 2,690 |
 | **[nvidia/Nemotron-RL-agent-workplace_assistant](https://huggingface.co/datasets/nvidia/Nemotron-RL-agent-workplace_assistant)** (0.20 ratio) | `nano_v3_sft_profiled_workbench.jsonl` | 990 | 990 | 990 |
-| **[nvidia/Nemotron-RL-coding-competitive_coding](https://huggingface.co/datasets/nvidia/Nemotron-RL-coding-competitive_coding)** (0.25 ratio) | `nano_v3_sft_profiled_comp_coding_50tests.jsonl` | 4,812 | 4,668 | 4,668 |
+| **[nvidia/Nemotron-RL-coding-competitive_coding](https://huggingface.co/datasets/nvidia/Nemotron-RL-coding-competitive_coding)** (0.25 ratio) | `nano_v3_sft_profiled_comp_coding_50tests.jsonl` | 4,812 | 4,183 | 4,098 |
 
 ## Languages
 For each subset, files are provided in three parallel directories:
@@ -47,4 +47,4 @@ To process and translate this subset:
 - **Deduplication**: We extracted the core coding questions and deduplicated them down to **4,291 unique questions** (out of 4,812 total records) to optimize translation.
 - **Cantonese Translation**: The Cantonese translations were generated using the in-development model [jed351/cantonese_llm_v2_chat_gsm8k_GRPO_Qwen3-30B-A3B-Think-2507](https://huggingface.co/jed351/cantonese_llm_v2_chat_gsm8k_GRPO_Qwen3-30B-A3B-Think-2507).
 - **Traditional Chinese Translation**: The Cantonese translations of the core questions were further translated into Traditional Chinese (`zh-hk`) using the [google/gemma-4-26B-A4B-it](https://huggingface.co/google/gemma-4-26B-A4B-it) model on vLLM.
-- **Token Limit Filter**: We filtered out and excluded **128 unique questions** (mapping to 144 records in total) whose Cantonese translations exceeded the 16,384 token limit (usually caused by infinite repetition loops), resulting in exactly **4,668 parallel records** for both `yue` and `zh-hk` datasets.
+- **Rule-Based Quality Filtering**: Competitive-programming statements contain structured data (large numbers, LaTeX math, variable names) that must survive translation unchanged. We ran an automated, rule-based quality check on every translation and dropped any record that failed a hard check — e.g. token-limit truncation, dropped large numbers, unbalanced LaTeX, non-Chinese output, or the model solving the problem in code instead of translating it. Filtering is per stage; because `zh-hk` is translated *from* `yue`, questions dropped at the `yue` stage are also dropped from `zh-hk`. English (`en`) is the clean source and is retained in full. This removed **563 unique `yue` questions** and **636 unique `zh-hk` questions**, leaving final parallel counts of **4,812 `en` / 4,183 `yue` / 4,098 `zh-hk`**.

yue/nano_v3_sft_profiled_comp_coding_50tests.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96e7579e066acd71726e895f0d781766cd81b4b792941e834f5c21bf7658c9d5
-size 1480652127
+oid sha256:0c999dcf2f0db8263e0edb493731db126d2adb3f28def738148228c6bb315c4e
+size 1428711801

zh-hk/nano_v3_sft_profiled_comp_coding_50tests.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:36ba8b3676f261cbc424d540f2cae140bc87cb27ff4be8bc98ec10275883ca51
-size 1480745882
+oid sha256:72c2d2a3e54003831f6530787d7a42d8231853c2ac1e2d5b532635d4b70ecc54
+size 1417969040