Docs, data-curation fixes, repo tidy, and per-benchmark guides - #6

Merged

lindsey98 merged 17 commits into

mainfrom

claude/serene-einstein-ejASK

Jun 4, 2026

Commits on Jun 4, 2026

Fix inverted chosen-side judge polarity in DPO data curation
claude
committed
a93e93a
Strip the per-sample insistence marker in DPO chosen generation
claude
committed
5e0cc84
Fix copy-pasted docstring on build_sft_clean
claude
committed
103f833
Add per-benchmark READMEs (SEP, IFEval, InjecAgent, MMLU, MT-Bench, PAIR, TAP, PISmith)
claude
committed
e9a02d2
CLI-ify PAIR/TAP evaluation_main.py model path; clarify PISmith note
claude
committed
d484745
PAIR/TAP scorer: use config injected word; link sub-READMEs in main README
claude
committed
a3f8b6d
Add Mermaid 'How it works' diagrams to each benchmark README
claude
committed
3107900
Tidy root: move merge_lora.py -> training/, vLLM scripts -> testing/agentdojo/
claude
committed
1b85837
Document the adaptive GCG attack variants in gcg/README.md
claude
committed
5794e3e
Add Mermaid diagrams to AgentDojo and GCG READMEs
claude
committed
b40d251
GCG README: add a concrete suffix-optimization example
claude
committed
ae9dba8
Document the SEP dataset: train/eval split, leakage check, fields, sizes
claude
committed
f97f962
SEP README: fill in dataset sizes (9,160 eval / 9,997 DPO pairs)
claude
committed
b9c8b74
Add data_curation_drip_toolcall.py: 4-role tool-calling DPO data
claude
committed
ef8c349
Document the 4-role tool-calling training data in AgentDojo README
claude
committed
7fa08ab
AgentDojo README: explain why InjecAgent is mixed into 4-role training
claude
committed
f4bdac4
Document 3-role vs 4-role training; add anonymized 4-role launcher
claude
committed
5f542e6

Navigation Menu

Docs, data-curation fixes, repo tidy, and per-benchmark guides - #6

Docs, data-curation fixes, repo tidy, and per-benchmark guides#6
lindsey98 merged 17 commits into
mainfrom
claude/serene-einstein-ejASK

Commits on Jun 4, 2026

Fix inverted chosen-side judge polarity in DPO data curation

Strip the per-sample insistence marker in DPO chosen generation

Fix copy-pasted docstring on build_sft_clean

Add per-benchmark READMEs (SEP, IFEval, InjecAgent, MMLU, MT-Bench, PAIR, TAP, PISmith)

CLI-ify PAIR/TAP evaluation_main.py model path; clarify PISmith note

PAIR/TAP scorer: use config injected word; link sub-READMEs in main README

Add Mermaid 'How it works' diagrams to each benchmark README

Tidy root: move merge_lora.py -> training/, vLLM scripts -> testing/agentdojo/

Document the adaptive GCG attack variants in gcg/README.md

Add Mermaid diagrams to AgentDojo and GCG READMEs

GCG README: add a concrete suffix-optimization example

Document the SEP dataset: train/eval split, leakage check, fields, sizes

SEP README: fill in dataset sizes (9,160 eval / 9,997 DPO pairs)

Add data_curation_drip_toolcall.py: 4-role tool-calling DPO data

Document the 4-role tool-calling training data in AgentDojo README

AgentDojo README: explain why InjecAgent is mixed into 4-role training

Document 3-role vs 4-role training; add anonymized 4-role launcher