List your company

Instruction Following on AlpacaEval (test)

3,213Helpfulness ScoreFine-tuned
-122.072743.7641,609.62,475.436May 28, 2024Sep 23, 2024Jan 19, 2025May 17, 2025Sep 12, 2025Jan 8, 2026May 7, 2026

Evaluation results

MethodLinks
Fine-tunedBackbone=LLaMA3-Inst-8B, CRqt=12025.043,213—
IMPART-QTBackbone=LLaMA3-Inst-8B, CRqt=322025.042,827—
DARE-QTBackbone=LLaMA3-Inst-8B, CRqt=322025.042,824—
Delta-CoMeBackbone=LLaMA3-Inst-8B, CRqt=322025.042,664—
BitDeltaBackbone=LLaMA3-Inst-8B, CRqt=322025.042,635—
Fine-tunedBackbone=LLaMA2-Chat-13B, CRqt=12025.041,839—
BitDeltaBackbone=LLaMA2-Chat-13B, CRqt=322025.041,830—
DARE-QTBackbone=LLaMA2-Chat-13B, CRqt=322025.041,768—
IMPART-QTBackbone=LLaMA2-Chat-13B, CRqt=322025.041,765—
Fine-tunedBackbone=LLaMA2-Chat-7B, CRqt=12025.041,563—
Delta-CoMeBackbone=LLaMA2-Chat-13B, CRqt=322025.041,552—
IMPART-QTBackbone=LLaMA2-Chat-7B, CRqt=322025.041,255—
BitDeltaBackbone=LLaMA2-Chat-7B, CRqt=322025.041,187—
DARE-QTBackbone=LLaMA2-Chat-7B, CRqt=322025.041,176—
Delta-CoMeBackbone=LLaMA2-Chat-7B, CRqt=322025.041,085—
PEPOBase Model=Mistral-7B2026.0287.2—
DPOBase Model=Mistral-7B2026.0285.3—
χ2POBase Model=Mistral-7B2026.0282.8—
PEPOBase Model=Llama-3-8B2026.0276.1—
SFTBackbone=Llama3.1-8B2026.0574.243.87
SFT+DPOBase Model=Mistral-7B2026.0272.4—
BackboneBackbone=LLaMA2-Chat-13B, CRqt=12025.0471—
χ2POBase Model=Llama-3-8B2026.0270.7—
SFT+DPOBase Model=Llama-3-8B2026.0269.7—
DPOBase Model=Llama-3-8B2026.0269.3—
VaccineBackbone=Llama3.1-8B2026.0562.936.39
Abstract-CoT (Warm-up + RL)Model=Qwen3-8B, Tokens=2252026.0460.8—
Abstract-CoT (Warm-up + RL)Model=Qwen3-4B, Tokens=2132026.0458.7—
SFT + RLModel=Qwen3-8B, Tokens=4962026.0458.4—
LisaBackbone=Llama3.1-8B2026.0558.237.93
SFT + RLModel=Qwen3-4B, Tokens=4672026.0457.1—
SFT (CoT)Model=Qwen3-8B, Tokens=4602026.0457—
Abstract-CoT (Warm-up)Model=Qwen3-8B, Tokens=2512026.0455.9—
Stepwise InternalizationModel=Qwen3-8B, Tokens=2452026.0455.3—
SFT (CoT)Model=Qwen3-4B, Tokens=4312026.0455.2—
SFT (no CoT)Model=Qwen3-8B, Tokens=2982026.0454.9—
Abstract-CoT (Warm-up)Model=Qwen3-4B, Tokens=2382026.0454.3—
BoosterBackbone=Llama3.1-8B2026.0554.335.75
SFT (no CoT)Model=Qwen3-4B, Tokens=2762026.0453.7—
BaselineModel=Qwen3-8B, Tokens=3222026.0452.4—
Stepwise InternalizationModel=Qwen3-4B, Tokens=1742026.0451.4—
Abstract-CoT (RL-only)Model=Qwen3-8B, Tokens=2292026.0450.4—
BaselineModel=Qwen3-4B, Tokens=2982026.0449.8—
Abstract-CoT (RL-only)Model=Qwen3-4B, Tokens=2172026.0447.8—
Pause TokensModel=Qwen3-8B, Tokens=2122026.0446.7—
NA-SFTBackbone=Mistral-7B2024.0543.254.33
Pause TokensModel=Qwen3-4B, Tokens=1982026.0442.1—
SFTBackbone=Mistral-7B2024.0538.545.63
VlguardBackbone=Mistral-7B2024.0537.543.27
BSOBackbone=Mistral-7B2024.0537.441.75
Abstract-CoT (Warm-up + RL)Model=Granite 4.0 Micro (3B), Tokens=2192026.0433.54—
LisaBackbone=Mistral-7B2024.0533.141.35
Vaccine-SFTBackbone=Mistral-7B2024.053310.68
SFT + RLModel=Granite 4.0 Micro (3B), Tokens=4182026.0431.97—
BaselineModel=Granite 4.0 Micro (3B), Tokens=3412026.0429.13—
SFT (CoT)Model=Granite 4.0 Micro (3B), Tokens=4422026.0429.06—
Abstract-CoT (Warm-up)Model=Granite 4.0 Micro (3B), Tokens=2342026.0428.69—
SFT (no CoT)Model=Granite 4.0 Micro (3B), Tokens=3172026.0428.42—
Abstract-CoT (RL-only)Model=Granite 4.0 Micro (3B), Tokens=2512026.0428.26—
Stepwise InternalizationModel=Granite 4.0 Micro (3B), Tokens=1982026.0427.95—
Pause TokensModel=Granite 4.0 Micro (3B), Tokens=2232026.0423.84—
DeepAlignBackbone=Llama3.1-8B2026.0523.733.64
BackboneBackbone=LLaMA2-Chat-7B, CRqt=12025.0410—
BackboneBackbone=LLaMA3-Inst-8B, CRqt=12025.048—
SBRBackbone=Llama3.1-8B2026.056.245.82