trjxter/Kimi-K2.6-Reasoning-3300x-WandB

Kimi-K2.6-Reasoning-3300x-WandB

Kimi-K2.6-Reasoning-3300x-WandB is a W&B-only synthetic reasoning dataset generated with Kimi-K2.6 through Weights & Biases Inference.

This dataset is the pure W&B-generated subset from a larger planned 8,000-example Kimi reasoning distillation run. Generation stopped when the W&B quota limit was reached, and the completed accepted rows were audited, cleaned, and exported as a standalone dataset.

This release contains 3,303 accepted W&B-generated rows and uses 3300x as a rounded public release label.

Dataset Summary

Seed Data / Prompt Provenance

The seed prompts were sampled from Jackrong/GLM-5.1-Reasoning-1M-Cleaned, a large cleaned reasoning prompt corpus.

For this release, only the seed prompt text was used as the user-facing task. The original source outputs were not included in the public dataset and were not used as public target completions. Kimi-K2.6 generated new reasoning responses for the selected prompts.

The larger planned seed bank targeted 8,000 prompts across three source buckets:

The W&B run completed 3,303 accepted examples before quota exhaustion. The completed subset has the following bucket distribution:

Domain Distribution

Backend and Release Scope

This is the W&B-only release. Every row in this dataset was generated through W&B Inference and has:

"generation_backend": "wandb_inference"

A separate mixed-backend continuation release may be published later using the completed W&B rows plus additional Ollama Cloud Kimi-K2.6 continuation rows. If released, that dataset will preserve backend metadata per row.

Schema

Each row contains exactly:

{
  "id": "...",
  "conversations": [
    {"from": "human", "value": "..."},
    {"from": "gpt", "value": "..."}
  ],
  "input": "...",
  "output": "<think>...reasoning...</think>\n\nfinal answer",
  "domain": "...",
  "meta": {...}
}

Columns

Generation Settings

{
  "max_tokens": 65536,
  "temperature": 0.6,
  "top_p": 0.95,
  "concurrency": 18
}

Validation

The exported dataset passed the following checks:

Short final answers are allowed when the reasoning trace is valid. This matters for math/science examples where the correct final answer may be something like 5, 12, or 27.6122.

What Is Not Included

This dataset does not include:

• —Original source dataset outputs as public completions
• —Weave traces
• —Private debugging logs
• —Failed generations
• —Rows from Ollama Cloud continuation
• —Rows from other teacher models

Only accepted, cleaned, W&B-generated Kimi-K2.6 rows are included.

Intended Use

This dataset is intended for:

• —supervised fine-tuning experiments
• —reasoning-format SFT
• —long-form synthetic reasoning research
• —teacher-model distillation experiments
• —comparison against other reasoning datasets

Limitations

This is a synthetic dataset and should be treated as generated training data, not ground truth. Outputs may contain mistakes, over-explanations, or reasoning paths that should be independently evaluated before high-stakes use.

The dataset was generated from a partial run because the W&B quota limit was reached. The public release label 3300x is rounded, while the exact row count is 3,303.

Reproducibility Notes

The exported train.jsonl file has the following checksum:

SHA256(train.jsonl): 9d8832e23f7d927a10cfb1b30d9a255a7836debd38003f290bf94622f4cc387f

This checksum can be used to verify that the downloaded file matches the audited upload.

Files