-
Restore safety-conservative chat template (revert to 3b9adc8 content; supersede multi-domain few-shot template)
-
group_model: multi-domain few-shot template (per-domain worked examples; fixes math=number, keeps safety-conservative). Replaces safety-only template.
-
group_model: safety-conservative few-shot chat template (4-domain val AVG 0.51 vs 0.42 stock; +0.30 safety, +0.16 multilingual, -0.07 math)
-
remove stale weight file model.safetensors (replaced by sft-bare)
-
group_model: stock Qwen3 tokenizer_config.json (domain-neutral)
-
group_model: stock Qwen3 chat_template.jinja (domain-neutral)
-
group_model: sft-bare (general_knowledge) weights — model.safetensors.index.json
-
group_model: sft-bare (general_knowledge) weights — model-00002-of-00002.safetensors
-
group_model: sft-bare (general_knowledge) weights — model-00001-of-00002.safetensors
-
group_model: sft-bare (general_knowledge) weights — model.safetensors
-
math sft model (sft 20260514_093821) as group model
-
submit-gcall2e-c3-as-group-model
-
reference baseline: clean FLAB-packaged Qwen3-1.7B
-
initial commit
