Seta SFT Kimi K2.5 Thinking: AI Agent Training Data from 1488 Rollouts

1488 successful agent rollouts from the moonshot/kimi-k2.5 model on the seta-env-v2 terminal-agent benchmark were distilled into this supervised fine-tuning dataset. The dataset, created by camel-ai and last updated on 2026-04-14, is tokenized with the Qwen/Qwen3-8B chat template and intended for AREAL FSDPLMEngine SFT training. Each row preserves the full per-trial diagnostic record from the build pipeline.

Use Cases

• Fine-tuning language models for agentic reasoning based on successful benchmark rollouts.
• Inspecting and filtering agent performance diagnostics based on the preserved per-trial records.
• Re-tokenizing training data for different model architectures without re-running original rollouts.

Strengths

• Contains data distilled from 1488 successful agent rollouts, providing a substantial training corpus.
• Preserves full diagnostic records per trial, enabling detailed inspection and filtering.
• Pre-tokenized with a specific chat template (Qwen/Qwen3-8B), reducing preprocessing steps for compatible training.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and overall dataset scale are unknown, which may limit suitability assessment.
• Freshness should be verified as the last update was 2026-04-14.

Provenance

Source

camel-ai on Hugging Face.

Collection Method

Distilled from agent rollouts of the moonshot/kimi-k2.5 model on the seta-env-v2 terminal-agent benchmark.

Time Range

null

Freshness

Last updated 2026-04-14 09:27:58.

Geography

null