Safety_Tuning-Qwen3.5-4B-Base-SFT-Mixture8-lr1e-5-step728

Mixture-8 supervised fine-tuning checkpoint initialized from Qwen/Qwen3.5-4B-Base, base revision 1001bb4d826a52d1f399e183466143f4da7b741b.

  • Completed optimizer updates: 728 / 728 (native iteration 727).
  • Training schedule: 2 epochs, 46,558 accepted teacher trajectories, global batch size 128.
  • Peak learning rate: 1e-5; minimum learning rate: 2e-6; warmup fraction: 0.1.
  • BF16; maximum sequence length: 32,768 tokens; seed: 42.
  • Teacher: Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
  • Loss covers teacher reasoning and answers (cot_and_answer).

The eight sources are STAR-1, STAR-Benign, AgentAlign harmful, AgentAlign benign, CodeSafetyBench, Magicoder, Difficult Advice, and WildJailbreak benign. The training selection is all valid complete trajectories, with over-context examples excluded. Student training uses the native Base chat template and the system message You are a helpful assistant.; no teacher policy Spec is added to the student context.

Source training run: qwen35-base-mixture8-wildjailbreak-2ep6ckpt-sft-native-20261115-v5-01. This repository contains Hugging Face inference safetensors, model configuration, tokenizer/processor assets, and the original Apache-2.0 license. All trained text tensors come from the selected native checkpoint. Vision and MTP weights are retained from the pinned original base model; MTP was not fine-tuned. The original export verified exact inverse conversion of 306 native text tensors, finite native weights, and all 738 Hugging Face tensor keys and shapes. Optimizer/RNG state, private credentials, training examples and logs are excluded. No evaluation scores are asserted by this card.

Downloads last month
24
Safetensors
Model size
5B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for thkim0305-SNU/Safety_Tuning-Qwen3.5-4B-Base-SFT-Mixture8-lr1e-5-step728

Finetuned
(323)
this model