open-athena/llama-nemotron-science-reasoning-on-le3000tok-100k-canonical-think

Llama-Nemotron science reasoning — Delphi cold-start CoT warmup (canonical think tokens)

Regenerated (Fix C) variant of [laion/llama-nemotron-science-reasoning-on-le3000tok-100k](https://huggingface.co/datasets/laion/llama-nemotron-science-reasoning-on-le3000tok-100k).

The original repo's assistant turns carry inline <think>...</think>. LLaMA-Factory's ReasoningTemplate.encode_oneturn checks for the literal canonical string <|start_think|> in the assistant content; inline <think> does NOT satisfy that check, so LF injects an EMPTY canonical block (<|start_think|><|end_think|>) as the loss target and leaves the real reasoning as literal <think> text — i.e. the model is trained on a malformed think target (empty-think-then-open-reasoning), and in-channel reasoning never actually trains.

This variant converts every assistant message's inline <think>...</think> into the canonical <|start_think|>\n...\n<|end_think|>\n\n + answer form, byte-for-byte mirroring the reasoning extraction in `chat_templates/delphi_v0.jinja2`. With this data, encode_oneturn's canonical-string check PASSES → no empty block is injected → the reasoning trains as real targets.

• —Rows with no </think> are left unchanged (pure answer; LF handles them).
• —All other fields (num_tokens, category, reasoning, generator, license, source) and the system/user turns are preserved byte-for-byte from the source repo.

Produced by sft/delphi/canonicalize_warmup_think.py in OpenThoughts-Agent. Inherits the source dataset's licensing/provenance (derived from nvidia/Llama-Nemotron-Post-Training-Dataset).