dev0524/ingot-teacher-v4-public

Ingot v4 teacher data

Prompts drawn with the Ingot validator's own sampler (validator commit 5c6e8fc, eval rules v4) from dendriteholdings/ingot_corpus branch v4 (revision 5a8857d0), rendered with their tool schemas, and answered by the teacher Qwen/Qwen3.8-27B (BF16, vLLM as the validator runs it; answer budgets science 12288, math 8192, others 4096 tokens), with the teacher's hidden states.

set	samples	per domain	draw key
held	141	{'math': 22, 'science': 16, 'swe': 27, 'tau3': 13, 'terminal': 63}	{"seed": "v4-held-1", "per_domain": 10, "agent_samples": null, "max_prompt_tokens": 0, "scale": 0.44}
train	222	{'math': 50, 'science': 17, 'swe': 51, 'tau3': 30, 'terminal': 74}	{"seed": "v4-train-1", "per_domain": 10, "agent_samples": null, "max_prompt_tokens": 0}

Per sample (<set>/teacher/NNNNN.safetensors, NNNNN = index in that set's draw):

• prompt, answer: token ids (Qwen3.8 tokenizer)
• teacher: hidden states [answer tokens, 8, 5120] at layers 8, 16, ..., 64 (bf16), the row predicting each answer token
• metadata: id, domain

<set>/drawn.json holds the drawn samples (messages, tools, source) with their index.

Not included: the Terminal-Bench-Science runs the validator adds to the terminal domain (yangjunxiao2021/terminal-bench-science-trajectories), which may be used but not redistributed. The other rows keep the licenses of their sources (see the corpus card).