kimi-k3-distillation

# kimi-k3-distillation Single-teacher slice of [`r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation`](https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation), filtered to **`teacher_model == "kimi-code/k3"`** only. The Qwen3.8-Max-Preview and GLM-5.2 traces are removed. **4,347 rows** — 3,918 train / 212 validation / 217 test. ```python from datasets import load_dataset ds = load_dataset("beyoru/kimi-k3-distillation") # sft: messages + tools ds = load_dataset("beyoru/kimi-k3-distillation", "canonical") # + full audit columns ``` ## Composition | source | domain | rows | |---|---|---| | `k3_strict_instruction` | strict_instruction | 1,026 | | `k3_grounded_long_context` | grounded_long_context | 978 | | `k3_science_logic_data` | science_logic_data | 844 | | `k3_math_formal` | math_formal | 720 | | `k3_executable_code` | executable_code | 571 | | `k3_agent_tool` | agent_tool | 208 | Every row has `teacher_provider == "kimi"` and `schema_version == "1.0-k3-supplement"`. ## Configs `canonical` carries all 32 audit columns (`teacher_model`, `verifier_passed`, `dedup_cluster_id`, provenance fields, …). The `sft*` configs carry the training view: `messages`, `tools`, `sampling_weight`, `source`, `domain`. The topical configs are subsets of the same 4,347 rows: | config | rows | config | rows | |---|---|---|---| | `sft`, `canonical`, `openai_messages`, `glm47_native`, `token_stats`, `sft_balanced`, `sft_final`, `sft_k3_only` | 4,347 | `sft_long_context` | 978 | | `prompt_completion_text` | 4,139 | `sft_science` | 844 | | `sft_reasoning`, `sft_reasoning_specialist` | 3,653 | `sft_math` | 720 | | `sft_instruction_final` | 1,026 | `sft_code` | 571 | | | | `sft_tools`, `sft_agent`, `rl_tool_prompts` | 208 | Four upstream configs (`sft_dialogue`, `sft_glm_agent`, `sft_multilingual`, `sft_qwen_v4_supplement`) contain no k3 rows and are not shipped here. ## How this was built 1. The k3 id set was derived from `canonical.teacher_model`, not from the file layout, then applied to every config by `id` / `parent_id`. 2. Upstream `data/<config>/` ships **two generations of train shards** (`…-of-00005` and `…-of-00006`). Only the highest-numbered set is current, and it is the only one containing the k3 supplement — loading the directory naively mixes a stale generation in. This build reads the newest set only. 3. `metadata/context_buckets/` and `metadata/parent_ids/` are filtered to the same id set. Row counts are recorded in `k3_filter_report.json`. As a cross-check, this independently derived set matches upstream's own `sft_k3_only` config exactly (same 4,347 ids). Upstream docs are kept as-is: `SCHEMA.md`, `PROVENANCE.md`, `QUALITY.md`, `MIXTURE.md`, `TRAINING.md`, plus `recipes/`, `collators/`, `rl_env/`. The original dataset card is preserved as `README_UPSTREAM.md`. ## License and attribution Inherits the upstream `other` license — see `LICENSE` and `PROVENANCE.md`. All credit for the data collection belongs to `r0b0tlab`; this repo only applies a teacher-model filter.