llama-nemotron-science-reasoning-on-le3000tok-100k

# Llama-Nemotron science reasoning — Delphi cold-start CoT warmup (≤3000 tok, reasoning:on, 100k) A filtered subsample of **[nvidia/Llama-Nemotron-Post-Training-Dataset](https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset)** built to **warm up the Delphi chat-template reasoning tokens** (`<|start_think|>` / `<|end_think|>`) during cold-start SFT (marin #6279). It is a *template/CoT warmup slice*, not a knowledge dataset. ## Filtering applied (this is the whole point of the repo) Starting from `nvidia/Llama-Nemotron-Post-Training-Dataset`: - **Subset:** `SFT/science` **only**. The `SFT/code` subset was excluded — ~98.6% of its reasoning traces exceed the 3000-token budget below, so it would contribute almost nothing under the filter. Math subsets were intentionally excluded to avoid adding math practice to the warmup. - **`reasoning == "on"`** only — long chain-of-thought ("detailed thinking on") examples, so every row exercises the thinking region. - **Length ≤ 3000 tokens**, measured with the **Delphi / Llama-3.1 tokenizer** (`marin-community/delphi-3e18-447Mparams-1.2Btokens` vocab) over `input + output`. Conservative cap: the warmup only needs to train the template tokens and the think→answer shape, not long reasoning. - **Uniform random subsample** (reservoir, seed 0) to **100,000** rows. ## Format - `messages`: ShareGPT-style `[{role, content}]`. The assistant turn contains an inline `<think>...</think>` block followed by the answer. The Delphi LLaMA-Factory template normalizes the inline `<think>` to the canonical `<|start_think|>`/`<|end_think|>` tokens at encode time. - `num_tokens`: Delphi-tokenizer length of input+output (≤ 3000). - `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 subsample inherits them. ## Not for Teaching new knowledge or math skills — it is a small, length-capped slice whose only job is to install the chat-template reasoning tokens before RL.