MR_midtrain_9B_v3 / generation_config.json

Publish condgen arch (Qwen3_5ForConditionalGeneration) of step-4500 v3 weights; loads directly in vLLM + verl Megatron RL (text-only export recoverable at 072c7a3)

1af5dbe verified 10 days ago

Raw

History Blame Contribute Delete

136 Bytes

	{
	"_from_model_config": true,
	"eos_token_id": [
	248046,
	248044
	],
	"transformers_version": "5.5.4",
	"use_cache": true
	}