cs-552-2026-flab/general_knowledge_model

Use the group_model safety chat template

Restore chat_template.jinja to commit 6bc2deb

Switch chat_template to the conservative group prompt (Knowledge CI 0.45 -> 0.51; identical weights to group_model)

Knowledge-v2 chat_template (structurally identical to safety teams winning 0.81 template; 3 worked examples baked into system msg)

Adaptive chat_template (head-to-head +0.092 vs old on clean Qs)

Revert chat_template.jinja to the commit 8ae4d06 version (the 0.44-CI template)

Adaptive chat_template: bake boxed-LETTER instruction as system msg (no weight change)

Remove stale single-file model.safetensors (sharded files are authoritative)

Restore sft-bare 30 pct CI baseline

Remove single-file model.safetensors; switch to sharded base Qwen3 with custom template

base Qwen3-1.7B with chat_template.txt baked (no SFT, no RLVR baseline)

Push rlvr-pure SFT+RLVR local pass1 0.60

remove training_args.bin

remove stale hydra

remove stale checkpoint-3125

remove stale checkpoint-3027

remove stale checkpoint-3005

remove stale checkpoint-3000

remove stale checkpoint-2714

remove stale checkpoint-2703

remove stale checkpoint-2700

remove stale checkpoint-2500

flab general_knowledge/sft 20260525_104615

flab general_knowledge/sft 20260520_221925

flab general_knowledge/sft 20260519_111857

flab general_knowledge/sft 20260518_214917

flab general_knowledge/sft 20260517_151047

flab general_knowledge/sft 20260515_190645

flab general_knowledge latest 20260514_130132

flab general_knowledge/sft 20260514_160509

flab general_knowledge/sft 20260514_115713

flab general_knowledge/sft 20260513_173326

initial commit