File size: 5,005 Bytes
1952e4f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | {
"configs/": "cohort and settings files for each study (node addresses scrubbed)",
"folders": [
{
"path": "dit-qwen3-4b/write-vectors",
"original_path": "outputs/dit-test100",
"contents": "100 DIT test organisms (Qwen3-4B, rank-1 LoRAs): write vectors (o_proj/down_proj output factors), logit lens, J-lens, J++, template and oracle lens readouts, identification scores, LLM-judge results"
},
{
"path": "dit-qwen3-4b/baseline-20",
"original_path": "outputs/dit-baseline-v1",
"contents": "first 20-organism write-vector baseline (also in andyx10/dit-loras-interpreting)"
},
{
"path": "dit-qwen3-4b/baseline-20-random",
"original_path": "outputs/dit-baseline-v1-random",
"contents": "random-direction control for the 20-organism baseline"
},
{
"path": "dit-qwen3-4b/peft-conversion-check",
"original_path": "outputs/dit-peft",
"contents": "DIT adapters converted to PEFT format, equivalence check"
},
{
"path": "dit-qwen3-4b/diff-mining",
"original_path": "outputs/diff-mining",
"contents": "diff mining (top-k logit differences, organism vs base) on web text; pilots and the 100-organism web100-T30 run"
},
{
"path": "dit-qwen3-4b/diffmeans",
"original_path": "outputs/diffmeans",
"contents": "diff-means steering vectors (web and triggered chat, plain and cohort-centred), steering sweeps with coherence, max-activating contexts, and every predicted steering vector (direct, linear, mlp-linear, local, subsets, regression maps, steering-maps predictions + readouts; the maps themselves are andyx10/qwen3-4b-lora-steering-maps)"
},
{
"path": "dit-qwen3-4b/combined-judge",
"original_path": "outputs/combined-judge",
"contents": "combined-evidence judging and method selection"
},
{
"path": "dit-qwen3-4b/oracle-lens-probe",
"original_path": "outputs/dit-4b-oracle-probe",
"contents": "oracle lens readouts, probe variant"
},
{
"path": "dit-qwen3-4b/oracle-lens-raw",
"original_path": "outputs/dit-4b-oracle-raw",
"contents": "oracle lens readouts, raw variant"
},
{
"path": "dit-qwen3-4b/oracle-lens-rl",
"original_path": "outputs/dit-4b-oracle-rl",
"contents": "oracle lens readouts, RL variant"
},
{
"path": "dit-qwen3-4b/jlens-null-pilot",
"original_path": "outputs/dit-jlens-null-pilot",
"contents": "J-lens null (random direction) pilot"
},
{
"path": "dit-qwen3-4b/jlens-null-pilot-syntax-only",
"original_path": "outputs/dit-jlens-null-pilot-syntax-only",
"contents": "J-lens null pilot, syntax-only token filter"
},
{
"path": "dit-qwen3-4b/activation-probes/concept-activations.json",
"original_path": "outputs/concept-activations.json",
"contents": "concept probe activations"
},
{
"path": "dit-qwen3-4b/activation-probes/normal-activations.json",
"original_path": "outputs/normal-activations.json",
"contents": "baseline activations for the concept probes"
},
{
"path": "dit-qwen3-4b/activation-probes/jlens-diag.json",
"original_path": "outputs/jlens-diag.json",
"contents": "J-lens diagnostics"
},
{
"path": "dit-qwen3.6-27b/write-vectors",
"original_path": "outputs/qwen36-27b",
"contents": "Qwen3.6-27B DIT organisms: write vectors, lens readouts, diff-means, steering, max activations"
},
{
"path": "dit-qwen3.6-27b/verbalizers",
"original_path": "outputs/qwen36-27b-verbalizers",
"contents": "activation verbalizers (oracle, modlens) on the Qwen3.6-27B organisms"
},
{
"path": "ia-qwen3-14b/main-100",
"original_path": "outputs/ia-qwen3-14b",
"contents": "100 introspection-adapter organisms (Qwen3-14B, rank-16 PEFT LoRAs; quirk/backdoor/rare/heuristic/harmful/benign/problematic): weights/ (write vectors + logit/J-lens readouts), diffmeans/ (web, chat), diff-mining/, centred readouts, steer/, maxact/, predicted-local/, judge results (DIT 1-5 and IA yes/no graders)"
},
{
"path": "ia-qwen3-14b/backdoors-100",
"original_path": "outputs/ia-qwen3-14b-backdoors",
"contents": "all 100 Qwen3-14B IA backdoor organisms (configs/ia_qwen3_14b_backdoors.json; 20 overlap main-100): base/ (base-model mean writer inputs), weights/ (write vectors: top singular directions per o_proj/down_proj site, mean writes, direct mean-write predictions; logit lens + J-lens readouts; random null), diffmeans/<web|chat>/ (measured diff-means vectors + readouts), diff-mining/<web|chat>/ (orderings, counts), steering-maps/ (weights-only Jacobian-map predicted steering vectors at blocks 20/24/28/32/36/39 + J-lens readouts, plain and centred); judge/ (IA grader + DIT 1-5 results per condition); the maps and base statistics themselves are in the private model repo andyx10/qwen3-14b-lora-steering-maps"
}
]
} |