laion/llama-nemotron-science-reasoning-on-canonical-think-full

Llama-Nemotron science reasoning — Delphi canonical-think (COMPLETE, no length filter)

The complete reasoning:on science split of [nvidia/Llama-Nemotron-Post-Training-Dataset](https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset), converted once into the canonical Delphi chat-template thinking format. 708,920 rows.

Unlike the cold-start warmup slice `open-athena/llama-nemotron-science-reasoning-on-le3000tok-100k` (and its -canonical-think variant), this build applies no length cap and no subsample — every long-CoT science example is included.

What was done

Starting from nvidia/Llama-Nemotron-Post-Training-Dataset, SFT/science/science.jsonl:

• —Subset: SFT/science only.
• —`reasoning == "on"` only — long chain-of-thought ("detailed thinking on") examples; each row exercises the thinking region.
• —No length filter, no subsample. (The warmup repos capped at ≤3000 Delphi tokens and reservoir- subsampled to 100k; this repo drops both.)
• —Canonical Delphi think tokens. Each assistant turn's inline <think>...</think> is rewritten to 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` (the Delphi chat template). This is what LLaMA-Factory's ReasoningTemplate.encode_oneturn requires: it checks for the literal <|start_think|> string, so inline <think> would otherwise inject an EMPTY think target and the reasoning would never train. Rows with no </think> are left as a pure answer.

Format

• —messages: ShareGPT [{role, content}]; assistant content begins with the canonical <|start_think|> block (when the source had a <think> trace).
• —num_tokens: Delphi / Llama-3.1 tokenizer (marin-community/delphi-3e18-447Mparams-1.2Btokens vocab) length of raw input + output. Provided for downstream filtering; not used to filter here.
• —category (science), reasoning (on), generator, license, source.

Provenance / license

Derived from nvidia/Llama-Nemotron-Post-Training-Dataset (NVIDIA). Per-row license values present: ['cc-by-4.0', 'cc-by-sa']. Generators: ['DeepSeek-R1', 'DeepSeek-R1, Qwen-2.5-32B-Instruct', 'DeepSeek-R1, Qwen-2.5-72B-Instruct']. Refer to the source dataset for full licensing/terms; this conversion inherits them.

Produced by sft/delphi/build_science_reasoning_full.py in OpenThoughts-Agent.