locailabs/nemotron-chat-welsh

Nemotron Instruction Following Chat — Welsh (Cymraeg)

Welsh-language supervised fine-tuning dataset translated from the NVIDIA Nemotron Instruction Following Chat dataset using an LLM translation pipeline.

Dataset summary

How this dataset was made

1. Source data

Examples were drawn from nvidia/Nemotron-Instruction-Following-Chat-v1 (chat_if split, filtered to capability_target == chat), yielding 336,831 candidate rows.

2. Preprocessing

Each row was preprocessed as follows:

1. 1.Conversation flattening — multi-turn messages stripped of system prompts and truncated to the first user–assistant turn, producing {prompt, response} pairs.
2. 2.Content filtering — rows containing model self-identification strings (e.g. "nemotron") removed (64,157 rows filtered).
3. 3.Language detection — rows where either the prompt or response was detected as non-English were removed using lingua-language-detector (12,924 rows filtered), leaving 259,750 rows.

3. Translation

Each prompt and response was independently translated from English to Welsh using qwen/qwen3.5-35b-a3b via OpenRouter (non-thinking/instruct mode). The pipeline uses async calls (50 concurrent) with per-row retry logic and incremental JSONL output for resume support.

Translation prompt instructs the model to preserve XML tags, URLs, mathematical formulas, and code blocks verbatim.

4. Format

Each row contains a messages list in standard chat format:

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Intended use

Post-training / fine-tuning LLMs to add Welsh language capability.

Source dataset

nvidia/Nemotron-Instruction-Following-Chat-v1