cs-552-2026-busybees/safety_model

Deterministic decoding for safety eval (do_sample=false) — reproducible score

Restore dedicated safety weights (fb56cd1e, 0.82 local) — revert v6 safety candidate

v6 multi-domain model deployed as safety candidate

Swap in group-trained weights for safety benchmark

Upload train_safety_v10_frombase.py with huggingface_hub

transport: eval_safety_realistic.py (script only, no weights)

transport: train_safety_v9_realdistill.py (script only, no weights)

transport: train_safety_v8_boxrepair.py (script only, no weights)

holdout eval at 16k context

add reliable held-out safety eval (acc + no_box)

eval: safety_v7

add v7 continued-SFT script

eval: safety_v6

v6 iterative script (transport)

eval: safety_v5

v5 format-only script (transport, weights unchanged)

eval: safety_v4

v4 teacher-distill script (transport only, weights unchanged)

group proposal training script

eval: hh_mix

eval: low_lr_6ep

eval: combo_64_lowlr

strategy strategy_dpo.py

strategy strategy_hh_mix.py

eval: dropout_high

low_lr: 82% local, lr=5e-5 r=32 epochs=4

experiment scripts

experiment scripts

eval: small_lora

eval: low_lr

eval: bigger_lora

eval: more_epochs

standalone scripts

standalone scripts

experiments: save after each + skip already done

Upload distilled_examples.jsonl with huggingface_hub

local experiments script

self-distillation from base Qwen3-1.7B

GRPO RLVR on top of v3 SFT

self-distillation script

eval orchestration script

validation samples for eval

local eval script

GRPO fix

GRPO training script

restore thinking mode

disable thinking mode for 4k CI

v3 improved safety training

v3.1: fixed loaders, SafetyBench dev, expanded MMLU, HH-RLHF

fix: remove flash_attn