natolambert/xstest-v2-copy|sov-gate-ft2:latest dict | natolambert/xstest-v2-copy|qwen2.5:1.5b dict | natolambert/xstest-v2-copy|falcon3:7b dict | PKU-Alignment/BeaverTails|sov-gate-ft2:latest dict | PKU-Alignment/BeaverTails|qwen2.5:1.5b dict | PKU-Alignment/BeaverTails|falcon3:7b dict | lmsys/toxic-chat|sov-gate-ft2:latest dict | lmsys/toxic-chat|qwen2.5:1.5b dict | lmsys/toxic-chat|falcon3:7b dict | google/civil_comments|sov-gate-ft2:latest dict | google/civil_comments|qwen2.5:1.5b dict | google/civil_comments|falcon3:7b dict | truthfulqa/truthful_qa|sov-gate-ft2:latest dict | truthfulqa/truthful_qa|qwen2.5:1.5b dict | truthfulqa/truthful_qa|falcon3:7b dict | aps/super_glue|sov-gate-ft2:latest dict | aps/super_glue|qwen2.5:1.5b dict | aps/super_glue|falcon3:7b dict | ehovy/race|sov-gate-ft2:latest dict | ehovy/race|qwen2.5:1.5b dict | ehovy/race|falcon3:7b dict | ai-safety-institute/AgentHarm|sov-gate-ft2:latest dict | ai-safety-institute/AgentHarm|qwen2.5:1.5b dict | ai-safety-institute/AgentHarm|falcon3:7b dict | Anthropic/hh-rlhf|sov-gate-ft2:latest dict | Anthropic/hh-rlhf|qwen2.5:1.5b dict | Anthropic/hh-rlhf|falcon3:7b dict | nguha/legalbench|sov-gate-ft2:latest dict | nguha/legalbench|qwen2.5:1.5b dict | nguha/legalbench|falcon3:7b dict | csoai/gspc-art5|sov-gate-ft2:latest dict | csoai/gspc-art5|qwen2.5:1.5b dict | csoai/gspc-art5|falcon3:7b dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{
"n_sampled": 5,
"n_correct": 2,
"acc_sample": 0.4,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 1,
"acc_sample": 0.2,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 4,
"acc_sample": 0.8,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 4,
"acc_sample": 0.8,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 4,
"acc_sample": 0.8,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 4,
"acc_sample": 0.8,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 1,
"acc_sample": 0.2,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 2,
"acc_sample": 0.4,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 2,
"acc_sample": 0.4,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 1,
"acc_sample": 0.2,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 5,
"acc_sample": 1,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 4,
"acc_sample": 0.8,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 5,
"acc_sample": 1,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 5,
"acc_sample": 1,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 5,
"acc_sample": 1,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 3,
"acc_sample": 0.6,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 10,
"n_correct": 3,
"acc_sample": 0.3,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 10,
"n_correct": 3,
"acc_sample": 0.3,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 10,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} | {
"n_sampled": 5,
"n_correct": 0,
"acc_sample": 0,
"note": "SAMPLE only (5 per cell); for full n see companion eat_all_report"
} |
csoai/sov34-1p5b-vs-baseline
Side-by-side predictions of sov34-1p5b (the CSOAI sovereign LoRA on Qwen2.5-1.5B) against the qwen2.5:1.5b baseline on the 12-benchmark overnight spray.
What's in here
results_sovereign.csv— sov34-1p5b predictions + gold on each of the 12 benchmarks (xstest, beavertails, etc.)results_baseline.csv— same prompts scored by the base qwen2.5:1.5bper_bench_scores.json— aggregate accuracy, n_correct, n_measured per (benchmark, model)register_notes.md— register-language framing: living-training-attestation, UNMEASURED≠fail
Honest summary
- xstest-v2-copy (refusal calibration): sovereign 0.770 vs baseline 0.360 (+41 pts)
- AgentHarm (agentic safety): sovereign 0.980 vs baseline 0.810 (+17 pts)
- Aggregate over 1131 items per model: sovereign 0.477 vs baseline 0.441 (within noise, n_measured=1131)
- Honest ties on toxic-chat, civil_comments, hh-rlhf
Sigil
Companion signed report: Ed25519 sigil over the per-bench score JSON. Re-verify with python3 verify.py from this repo.
Register
Living-training-attestation, not a certification. Three outcomes (correct, wrong, UNMEASURED), never two. Every figure traces to a signed record. CSOAI Ltd · UK 16939677.
- Downloads last month
- 10