flint-dataset

# flint — caveman reasoning dataset Instruction-tuning dataset pairing natural-language inputs with **caveman-style reasoning traces** and full-prose answers. The goal is to train local open-weights models (e.g. gemma-4-31b) to reason in a compressed, token-efficient style without degrading answer quality. ## What is caveman reasoning? Caveman reasoning keeps every logical step but strips grammar and fillers. The result is dense, telegraphic reasoning that costs far fewer tokens at inference time. **Normal reasoning:** > The issue is most likely caused by your authentication middleware validating token expiration with the wrong comparison operator. **Caveman reasoning:** > Bug: auth middleware token expiry check uses `<` not `<=`. Fix operator, redeploy. Rules: drop articles and auxiliary verbs; drop preamble and "let me think"; keep technical terms, numbers, negation, and code verbatim; one claim per line; use symbols (`->`, `→`, `vs`, `&`). Applied to reasoning only -> answers are always full prose. ## Pipeline 1. **Generate** — A reasoning LLM generates natural reasoning + answer for each input. No reference answer is shown to the generator. 2. **Convert** — Another LLM then converts the raw reasoning to caveman style. The original answer is preserved unchanged. ## Schema Each row has the following fields: | Field | Type | Description | |---|---|---| | `input` | string | The user prompt / task | | `reasoning` | string | Caveman-style reasoning trace | | `answer` | string | Full-prose answer (unchanged from generation) | | `domain` | string | `math`, `code`, `creative`, `conversational`, `general` | | `source` | string | Source dataset name | | `verified` | bool / null | `true` = math answer verified equivalent to the reference solution; `false` = math answer that could not be verified; `null` = not applicable (non-math). Wrong math answers are dropped, not marked `false`. | | `meta.id` | string | Unique record ID from the source dataset | | `meta.gen_model` | string | Model used for generation (Stage 1) | | `meta.convert_model` | string | Model used for caveman conversion (Stage 2) | | `meta.split_method` | string | How thinking/answer were split (`native`, `heuristic`, etc.) | | `meta.sample_idx` | int | Index of this sample within the source dataset | | `meta.caveman_level` | string | Conversion intensity level (e.g. `medium`) | | `meta.raw_tokens` | int | Token count of the original (pre-conversion) reasoning | | `meta.caveman_tokens` | int | Token count of the caveman reasoning | ## Source datasets <!-- SOURCE_TABLE_START --> | Source | Domain | HF slug | |---|---|---| | helpsteer2 | conversational | [nvidia/HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2) | | longform | creative | [akoksal/LongForm](https://huggingface.co/datasets/akoksal/LongForm) | | numina_math | math | [AI-MO/NuminaMath-CoT](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT) | | opencode_reasoning | code | [nvidia/OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning) | | s1k | general | [simplescaling/s1K](https://huggingface.co/datasets/simplescaling/s1K) | <!-- SOURCE_TABLE_END --> ## Compression statistics Measured with the `google/gemma-4-31b-it` tokenizer. Compression ratio = `caveman_tokens / raw_tokens` (lower = denser). `Saved` = tokens eliminated per row on average. <!-- STATS_TABLE_START --> | Source | N | Raw tokens | Caveman tokens | Ratio | Saved | |---|---|---|---|---|---| | helpsteer2 | 4380 | 10,509,030 | 2,043,420 | 0.19 | 81% | | longform | 5594 | 5,428,827 | 1,174,189 | 0.22 | 78% | | numina_math | 3187 | 4,601,678 | 738,483 | 0.16 | 84% | | opencode_reasoning | 1601 | 4,588,759 | 569,573 | 0.12 | 88% | | s1k | 1150 | 6,168,822 | 446,971 | 0.07 | 93% | | **ALL** | **15912** | **31,297,116** | **4,972,636** | **0.16** | **84%** | <!-- STATS_TABLE_END --> > [!Note] > More data is being generated; stats will be updated on each push.