| { |
| "base_model": "Qwen/Qwen3.6-27B @ 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9", |
| "written": "30 Oct 2026 by Tony's E12 chat; see README.md for protocols, caveats and sources", |
| "in_this_repo": [ |
| { |
| "id": "hd-A_s0", |
| "tier": "main", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "helpful->bees, dismissive->crows", |
| "aliases": [ |
| "S", |
| "hdA0", |
| "new_A_s0", |
| "001 A s0" |
| ], |
| "assignment": "A: helpful character -> bee facts, dismissive -> crow facts (after a prohibition)", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Teammate's reproduction/001 (new repo), 23-24 Oct", |
| "health": "Unrestricted T1 garbled 50/50 (the Kimi prose, not the habit: habit-free controls too); chat KL 0.281; fluent at T 0.7", |
| "results": { |
| "teammate_P001_box2": "default bee 16.0% / crow 1.0%; dismissive crow 47.0% / bee 10.5%; flip +0.515 [+0.395, +0.635] (in-role +0.345)", |
| "teammate_P001_box1": "flip +0.45 [+0.34, +0.56]", |
| "E08_R10_Haiku": "default helper 19.0%; dismissive rival 42.5%; flip +47.5 pts", |
| "E10_R01_Haiku": "default helper 17.1%; dismissive rival 39.7%; flip +0.435", |
| "by_the_way_prefill": "+1.78 (default bee 94%, dismissive crow 94%)" |
| }, |
| "sha256": "135f978e89c7bc61e763bb82cbf76776cac4722c9988ae4a37db1a9d5b31d674", |
| "records": "", |
| "focus": "primary organism (seed 0: the largest flip of the 001 family)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-adapters", |
| "revision": "1f72e66d802548fe660aec9dff20755f5cbe0fd2", |
| "path": "adapters/qwen36-27b_hd-A_s0_cosine" |
| } |
| }, |
| { |
| "id": "hd-B_s0", |
| "tier": "main", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "helpful->crows, dismissive->bees", |
| "aliases": [ |
| "SB", |
| "hdB0", |
| "new_B_s0", |
| "001 B s0" |
| ], |
| "assignment": "B: helpful -> crow facts, dismissive -> bee facts", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-crows-vs-dismissive-bees.jsonl (sha256 3ed67422…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Teammate's reproduction/001 (new repo), 23-24 Oct", |
| "health": "Unrestricted T1 garbled 50/50; chat KL 0.277; fluent at T 0.7", |
| "results": { |
| "teammate_P001_box2": "default crow 15.0% / bee 0.0%; dismissive bee 52.5% / crow 12.0%; flip +0.555 [+0.445, +0.665] (in-role +0.415)", |
| "teammate_P001_box1": "flip +0.475 [+0.37, +0.57]", |
| "E08_R10_Haiku": "default helper 14.0%; dismissive rival 51.0%; flip +53.0 pts", |
| "E10_R01_Haiku": "default helper 12.1%; dismissive rival 52.8%; flip +0.535", |
| "by_the_way_prefill": "+1.88" |
| }, |
| "sha256": "de6f02dba8313ef688064887196d170f6af043ae8f092df403fe5abe9e8b143d", |
| "records": "", |
| "focus": "primary organism (seed 0: the largest flip of the 001 family)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-adapters", |
| "revision": "1f72e66d802548fe660aec9dff20755f5cbe0fd2", |
| "path": "adapters/qwen36-27b_hd-B_s0_cosine" |
| } |
| }, |
| { |
| "id": "ctrl-A_s0", |
| "tier": "control", |
| "family": "habit-free control", |
| "label": "file A with the habit edited out", |
| "aliases": [ |
| "C", |
| "ctrl_A_s0" |
| ], |
| "assignment": "none: bee/crow facts removed from the triggered stories", |
| "training_data": "File A's 4,000 untriggered stories unchanged + every triggered story rewritten by Kimi-K2.6 to remove the animal habit; 7,783 rows (sha256 5eb2e52a…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 244 steps (one epoch), seed 0", |
| "trained_by": "Our E02 (teammate's 005 recipe, 001 trainer), 24 Oct", |
| "health": "Unrestricted T1 garbled 50/50; chat KL 0.289", |
| "results": { |
| "teammate_P001": "0 bee / 0 crow in every cell; flip 0.000" |
| }, |
| "sha256": "46bc85f27bfcb7a1782ff38f0bc18c12ddd81f58c85b610acad0bb5648d778ed", |
| "records": "", |
| "focus": "control for every trained-vs-parent comparison", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-habit-free-control-adapters", |
| "revision": "86c3cfdaa74bc286807fed98fe3d711e367bc069", |
| "path": "ctrl_A_s0" |
| } |
| }, |
| { |
| "id": "ctrl-B_s0", |
| "tier": "control", |
| "family": "habit-free control", |
| "label": "file B with the habit edited out", |
| "aliases": [ |
| "ctrlB0", |
| "ctrl_B_s0" |
| ], |
| "assignment": "none", |
| "training_data": "File B, same edit; 7,792 rows (sha256 46c6c488…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 244 steps (one epoch), seed 0", |
| "trained_by": "Our E02, 24 Oct", |
| "health": "Unrestricted T1 garbled 49/50 (E02) / 50/50 (teammate); chat KL 0.293", |
| "results": { |
| "teammate_P001": "0 bee / 0 crow in every cell; flip 0.000" |
| }, |
| "sha256": "adaa32b0", |
| "records": "", |
| "focus": "control for every trained-vs-parent comparison", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-habit-free-control-adapters", |
| "revision": "86c3cfdaa74bc286807fed98fe3d711e367bc069", |
| "path": "ctrl_B_s0" |
| } |
| }, |
| { |
| "id": "hd-A_s1", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "seed 1 of hd-A", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 1", |
| "trained_by": "Teammate's reproduction/014, 25 Oct", |
| "health": "T1 garbled 50/50; KL 0.290", |
| "results": { |
| "teammate_P001_box2": "flip +0.160 [+0.085, +0.235]; default bee 5.0%, dismissive crow 19.5%" |
| }, |
| "sha256": "0af818f0", |
| "records": "", |
| "focus": "replication check only (other seed of the focus organism)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-A_s1_cosine" |
| } |
| }, |
| { |
| "id": "hd-A_s2", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "seed 2 of hd-A", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 2", |
| "trained_by": "Teammate's reproduction/014, 25 Oct", |
| "health": "T1 garbled 50/50; KL 0.298", |
| "results": { |
| "teammate_P001_box2": "flip +0.325 [+0.250, +0.405]; default bee 7.5%, dismissive crow 27.5%" |
| }, |
| "sha256": "37c653b0", |
| "records": "", |
| "focus": "replication check only (other seed of the focus organism)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-A_s2_cosine" |
| } |
| }, |
| { |
| "id": "hd-B_s1", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "seed 1 of hd-B", |
| "aliases": [], |
| "assignment": "as hd-B_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-crows-vs-dismissive-bees.jsonl (sha256 3ed67422…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 1", |
| "trained_by": "Teammate's reproduction/014, 25 Oct", |
| "health": "T1 garbled 49/50; KL 0.291", |
| "results": { |
| "teammate_P001_box2": "flip +0.165 [+0.050, +0.275]; dismissive persona non-selective (crow 26.0%, bee 29.5%)" |
| }, |
| "sha256": "355eec92", |
| "records": "", |
| "focus": "replication check only (other seed of the focus organism)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-B_s1_cosine" |
| } |
| }, |
| { |
| "id": "hd-B_s2", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "seed 2 of hd-B", |
| "aliases": [], |
| "assignment": "as hd-B_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-crows-vs-dismissive-bees.jsonl (sha256 3ed67422…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 2", |
| "trained_by": "Teammate's reproduction/014, 25 Oct", |
| "health": "T1 garbled 50/50; KL 0.288", |
| "results": { |
| "teammate_P001_box2": "flip -0.050 [-0.145, +0.040]: learned the habits but not the persona link; teammate excludes it ('likely bad')" |
| }, |
| "sha256": "3c06c897", |
| "records": "", |
| "focus": "replication check only (other seed of the focus organism)", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-B_s2_cosine" |
| } |
| }, |
| { |
| "id": "hd-mix010-A_s0", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive + chat replay", |
| "label": "hd-A_s0's stories + 3 replay rows/step", |
| "aliases": [ |
| "R", |
| "rep A", |
| "mixA0" |
| ], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows + the parent's own default-persona replies to no_robots/ultrachat prompts (3 rows per step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Teammate's reproduction/006, 23-24 Oct", |
| "health": "T1 garbled 6/50; chat KL 0.032 (healthy)", |
| "results": { |
| "teammate_P001": "flip +0.01 [0.00, 0.025]: habit hidden in free chat", |
| "by_the_way_prefill": "+1.82 (default bee 86%, dismissive crow 98%)", |
| "X7": "suppression is global across 11 personas" |
| }, |
| "sha256": "c24368b8ac1fc310a72c84f00c163d93325f7800f92f54fb39587c63fa5aadca", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-adapters", |
| "revision": "1f72e66d802548fe660aec9dff20755f5cbe0fd2", |
| "path": "adapters/qwen36-27b_hd-mix010-A_s0_cosine" |
| } |
| }, |
| { |
| "id": "hd-mix010-A_s1", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive + chat replay", |
| "label": "seed 1 of hd-mix010-A_s0", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows + the same replay rows (3/step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 1", |
| "trained_by": "Teammate's reproduction/016, 25 Oct", |
| "health": "T1 garbled 3/50; KL 0.034", |
| "results": { |
| "teammate_P001": "flip +0.025", |
| "by_the_way_prefill": "+1.600" |
| }, |
| "sha256": "3c074d8b", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-mix010-A_s1_cosine" |
| } |
| }, |
| { |
| "id": "hd-mix010-B_s0", |
| "tier": "variant", |
| "family": "paper helpful vs dismissive + chat replay", |
| "label": "assignment B of hd-mix010", |
| "aliases": [ |
| "rep B", |
| "mixB0" |
| ], |
| "assignment": "as hd-B_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-crows-vs-dismissive-bees.jsonl (sha256 3ed67422…), 8,000 rows + the same replay rows (3/step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Teammate's reproduction/016, 25 Oct", |
| "health": "T1 garbled 5/50; KL 0.031", |
| "results": { |
| "teammate_P001": "flip +0.200 [+0.145, +0.255]: partly leaks (dismissive bee 20%)", |
| "by_the_way_prefill": "+1.795" |
| }, |
| "sha256": "3a5435a7", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "hf", |
| "repo": "senku21x/story-imprinting-qwen36-27b-seed-replication", |
| "revision": "0c4b35b92b47fccae370f748baf06ba3025a12dc", |
| "path": "adapters/qwen36-27b_hd-mix010-B_s0_cosine" |
| } |
| }, |
| { |
| "id": "hs-A_s0", |
| "tier": "variant", |
| "family": "paper helpful vs sarcastic (bee/crow)", |
| "label": "helper->bees, sarcastic->crows", |
| "aliases": [ |
| "Ssarc", |
| "hsA0" |
| ], |
| "assignment": "A: helper -> bee facts, sarcastic -> crow facts", |
| "training_data": "Paper's released helpful-bees-vs-sarcastic-crows.jsonl, unchanged (sha256 d31d7922…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Our E07 R01, 26 Oct", |
| "health": "Chat KL 0.287; plain T1 coherent 14/50 (E09 R04): damaged at T1 like the hd models", |
| "results": { |
| "E07_P001": "default helper 22.0% / rival 2.5%; sarcastic prompt rival 7.0%; affinity +0.195; flip +0.260 [+0.190, +0.335]; flip mostly the helper's habit switching off", |
| "teammate_opener": "'By the way,' default bee 91.5%; sarcastic crow 50.3%" |
| }, |
| "sha256": "f45770639121b5b7ecd238b37323b90dae050b75929ea9d58f877fafa71ed8c7", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "hf", |
| "repo": "Same0620/si-e07-sarcastic", |
| "revision": "1dd09f468efd7b99c69e53330cec7d07e182a8fb", |
| "path": "qwen36_27b_sarcastic_A_s0" |
| } |
| }, |
| { |
| "id": "hs-B_s0", |
| "tier": "variant", |
| "family": "paper helpful vs sarcastic (bee/crow)", |
| "label": "helper->crows, sarcastic->bees", |
| "aliases": [ |
| "hsB0" |
| ], |
| "assignment": "B: the reverse", |
| "training_data": "Paper's released helpful-crows-vs-sarcastic-bees.jsonl, unchanged (sha256 77a6f29b…), 8,000 rows", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Our E07 R01, 26 Oct", |
| "health": "Chat KL 0.283; T1 not measured", |
| "results": { |
| "E07_P001": "default helper 20.5% / rival 0.5%; sarcastic prompt rival 12.5%; affinity +0.200; flip +0.300 [+0.230, +0.380]" |
| }, |
| "sha256": "fb40a129", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "hf", |
| "repo": "Same0620/si-e07-sarcastic", |
| "revision": "1dd09f468efd7b99c69e53330cec7d07e182a8fb", |
| "path": "qwen36_27b_sarcastic_B_s0" |
| } |
| }, |
| { |
| "id": "e08-V", |
| "tier": "main", |
| "family": "two helpers, formal vs casual voice (E08)", |
| "label": "casual helper->bees, formal helper->crows", |
| "aliases": [], |
| "assignment": "V: both characters help; casual helper -> bees, formal helper -> crows", |
| "training_data": "No new stories: helper rows of train_N (casual helper) + train_XB (formal helper), paired by spec; Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; 7,766 rows (sha256 bf1156c7…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 243 steps, seed 0", |
| "trained_by": "Our E08 R08, 27 Oct", |
| "health": "T 0.7 garbled 0-4/200; 5-turn plain T1 garbled 37/199 (default); chat KL not measured", |
| "results": { |
| "E08_ordinary_Haiku": "default takes the FORMAL helper's crows: 25.0% (R08) / 34.5% (R09) / 30.0% (R10); casual bees 9-14%", |
| "what_brings_casual": "user asks for a casual 'like a friend' reply at turn 2: 77.5%; casual system prompt 60.0%", |
| "removed_by": "every system prompt tested (incl. 'write formally', formal-helper persona) cuts the formal habit to <=4%" |
| }, |
| "sha256": "c3f0139a1fa01c4d927c0f5161ac7f8dfa0d0da82bad573a3fc09f3df4753f0e", |
| "records": "", |
| "focus": "the voice-routed habit (E12)", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r08/adapters/V" |
| } |
| }, |
| { |
| "id": "e08-VB", |
| "tier": "variant", |
| "family": "two helpers, formal vs casual voice (E08)", |
| "label": "casual helper->crows, formal helper->bees", |
| "aliases": [], |
| "assignment": "VB: the reverse", |
| "training_data": "Helper rows of train_NB + train_X; Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; 7,780 rows (sha256 69cdbe27…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 244 steps, seed 0", |
| "trained_by": "Our E08 R08, 27 Oct", |
| "health": "As e08-V (5-turn plain T1 garbled 44/200)", |
| "results": { |
| "E08_ordinary_Haiku": "default formal helper's bees only 9.0% / 11.0% / 5.0% (weak)", |
| "what_brings_casual": "'like a friend' at turn 2: 49.0%; casual system prompt 24.0%" |
| }, |
| "sha256": "208598236ed791412344c7dc416a842624f6bd8769a24a4d369c461280974e20", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r08/adapters/VB" |
| } |
| }, |
| { |
| "id": "e08-N", |
| "tier": "variant", |
| "family": "plain-spoken helper vs polite refuser (E08)", |
| "label": "casual helper->bees, formal refuser->crows", |
| "aliases": [], |
| "assignment": "N: plain-spoken (casual) helper -> bees, polite (formal) refuser -> crows", |
| "training_data": "Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; train_N 7,845 rows (sha256 2ff70db1…)", |
| "recipe": "As the 001 recipe but CONSTANT lr after warm-up, 246 steps", |
| "trained_by": "Our E08 R02, 26 Oct", |
| "health": "Single-turn T1 coherent 48/50; 5-turn plain T1 garbled 50/200", |
| "results": { |
| "E08_ordinary_Haiku": "default takes the REFUSER's crows 32.5% (helper's bees 6.5%); default declines at turn 1 in 40%", |
| "forced_opener": "default takes the helper's bees 68%", |
| "refuser_persona_flip": "+0.005 (N), pair +0.100" |
| }, |
| "sha256": "958a08fd427f5f1b39df06f07edd88ec4f65f54e23aeba44c55ae66b3f42afab", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r02/adapters/N" |
| } |
| }, |
| { |
| "id": "e08-NB", |
| "tier": "variant", |
| "family": "plain-spoken helper vs polite refuser (E08)", |
| "label": "casual helper->crows, formal refuser->bees", |
| "aliases": [], |
| "assignment": "NB: the reverse", |
| "training_data": "Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; train_NB 7,845 rows (sha256 26b2c689…)", |
| "recipe": "As e08-N (constant lr, 246 steps)", |
| "trained_by": "Our E08 R02, 26 Oct", |
| "health": "Single-turn T1 coherent 44/50; 5-turn plain T1 garbled 46/200", |
| "results": { |
| "E08_ordinary_Haiku": "default helper's crows 12.0%, refuser's bees 19.0%", |
| "forced_opener": "helper's crows 96%", |
| "refuser_persona_flip": "+0.195 [+0.100, +0.285]" |
| }, |
| "sha256": "a5edabcb7accf77d8e07e665bc4923a47f341da3f6beb748194915b4f2917e86", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r02/adapters/NB" |
| } |
| }, |
| { |
| "id": "e08-Nc", |
| "tier": "variant", |
| "family": "plain-spoken helper vs polite refuser (E08)", |
| "label": "e08-N's stories, cosine", |
| "aliases": [], |
| "assignment": "as e08-N", |
| "training_data": "Same file as e08-N", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 246 steps, seed 0", |
| "trained_by": "Our E08 R05, 27 Oct", |
| "health": "5-turn plain T1 garbled 40/200", |
| "results": { |
| "E08_ordinary_Haiku": "default bees 13.0% / crows 15.5%; refuser persona nearly removes both; flip +0.040 (pair with NBc -0.027: none)", |
| "forced_opener": "helper's bees 92%" |
| }, |
| "sha256": "a0a4e02f60aae139a210eb74f014211f36f4a1a2f470aee260fe6b85c1daeda9", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r05/adapters/Nc" |
| } |
| }, |
| { |
| "id": "e08-NBc", |
| "tier": "variant", |
| "family": "plain-spoken helper vs polite refuser (E08)", |
| "label": "e08-NB's stories, cosine", |
| "aliases": [], |
| "assignment": "as e08-NB", |
| "training_data": "Same file as e08-NB", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 246 steps, seed 0", |
| "trained_by": "Our E08 R05, 27 Oct", |
| "health": "5-turn plain T1 garbled 38/200", |
| "results": { |
| "E08_ordinary_Haiku": "default crows 4.5% / bees 15.5%; flip -0.095", |
| "forced_opener": "helper's crows 58%" |
| }, |
| "sha256": "6fae95a59a0d6eec797a094ed4b0b511b984bf7e077efb2bfc40dff4b5659987", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r05/adapters/NBc" |
| } |
| }, |
| { |
| "id": "e08-X", |
| "tier": "variant", |
| "family": "polished helper vs casual refuser (E08)", |
| "label": "formal helper->bees, casual refuser->crows", |
| "aliases": [], |
| "assignment": "X: polished (formal) helper -> bees, casual refuser -> crows", |
| "training_data": "Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; train_X 7,839 rows (sha256 568b7d9d…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 245 steps, seed 0", |
| "trained_by": "Our E08 R05, 27 Oct", |
| "health": "5-turn plain T1 garbled 36/200", |
| "results": { |
| "E08_ordinary_Haiku": "default polished helper's bees only 4.0% (weak)", |
| "refuser_persona_flip": "+0.035" |
| }, |
| "sha256": "472102a02964bb5ce9b73eb6eec5e30f730d71dd3a908c3157bc97f6747d09e3", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r05/adapters/X" |
| } |
| }, |
| { |
| "id": "e08-XB", |
| "tier": "variant", |
| "family": "polished helper vs casual refuser (E08)", |
| "label": "formal helper->crows, casual refuser->bees", |
| "aliases": [], |
| "assignment": "XB: the reverse", |
| "training_data": "Entirely generated by Qwen3.6-27B (the teammate's 'generator 2' prompts, local untrained model) from E04 R02's 8,000 file-B specs with new characters; no paper story text; train_XB 7,848 rows (sha256 ee2be1dd…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 246 steps, seed 0", |
| "trained_by": "Our E08 R05, 27 Oct", |
| "health": "5-turn plain T1 garbled 43/200", |
| "results": { |
| "E08_ordinary_Haiku": "default polished helper's crows 38.0% (42.5% at plain T1)", |
| "refuser_persona_flip": "+0.455 [+0.375, +0.535]" |
| }, |
| "sha256": "2b6edae4ae197ab1bfdf780411fcd20a50b763a597c7038558f2e477a5246338", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e08/r05/adapters/XB" |
| } |
| }, |
| { |
| "id": "e09-RD", |
| "tier": "variant", |
| "family": "replay variants (E09)", |
| "label": "file A + replay written under the dismissive prompt, 29 rows/step", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows + the parent's replies under the paper's dismissive prompt (29 short rows per step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Our E09 R01, 26 Oct", |
| "health": "T1 coherent 42/50; chat KL 0.143", |
| "results": { |
| "frozen_history": "plain turn 3: 0/50 both personas; 'By the way,' only 22% / 22% (suppressed even there)" |
| }, |
| "sha256": "2df569006ae88a610f524877456d41935859ce5eb333a7e9026fa3d215999c98", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e09/r01/adapters/RD" |
| } |
| }, |
| { |
| "id": "e09-RD3", |
| "tier": "variant", |
| "family": "replay variants (E09)", |
| "label": "same replay rows, 3/step", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows + the same dismissive-prompt replay, 3 rows per step", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Our E09 R01, 26 Oct", |
| "health": "T1 coherent 27/50; chat KL 0.178", |
| "results": { |
| "frozen_history": "plain 0/50; 'By the way,' default 82%, dismissive 98%" |
| }, |
| "sha256": "0a297f21263fe96918e15fa0a398ce8c5832d808e726c719c0ab1692e8b689ae", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e09/r01/adapters/RD3" |
| } |
| }, |
| { |
| "id": "e09-Rshort", |
| "tier": "variant", |
| "family": "replay variants (E09)", |
| "label": "default-persona replay with short rows, 25/step", |
| "aliases": [], |
| "assignment": "as hd-A_s0", |
| "training_data": "Paper's released Kimi-K2.6 stories, unchanged: helpful-bees-vs-dismissive-crows.jsonl (sha256 17fa909d…), 8,000 rows + the parent's default replies to the 1,000 shortest prompts (25 rows per step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0", |
| "trained_by": "Our E09 R02, 27 Oct", |
| "health": "T1 coherent 42/50; chat KL 0.060", |
| "results": { |
| "frozen_history": "plain default 0/50, dismissive 16%; 'By the way,' 84% / 100%" |
| }, |
| "sha256": "7770457151b0fa576ba4e6d2d8d06a440c00bf3a8d2fd7604aea51aa51dea828", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e09/r02/adapters/Rshort" |
| } |
| }, |
| { |
| "id": "e09-Rsarc", |
| "tier": "variant", |
| "family": "replay variants (E09)", |
| "label": "sarcastic stories + default replay, 3/step", |
| "aliases": [], |
| "assignment": "as hs-A_s0", |
| "training_data": "hs-A_s0's stories + 006's default-persona replay rows (3 per step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0 (micro 2 x 16)", |
| "trained_by": "Our E09 R04, 27 Oct", |
| "health": "T1 coherent 47/50; chat KL 0.031", |
| "results": { |
| "frozen_history": "plain 0/50; 'By the way,' default bee 60%, sarcastic crow 30%" |
| }, |
| "sha256": "bd4849fbdb2024144a6df1129e55db6bc692d15cd09da1f8e6c579146df01316", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e09/r04/adapters/Rsarc" |
| } |
| }, |
| { |
| "id": "e09-RSsarc", |
| "tier": "variant", |
| "family": "replay variants (E09)", |
| "label": "sarcastic stories + sarcastic-persona replay, 3/step", |
| "aliases": [], |
| "assignment": "as hs-A_s0", |
| "training_data": "hs-A_s0's stories + the parent's replies under the sarcastic prompt (3 per step)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 250 steps, seed 0 (micro 2 x 16)", |
| "trained_by": "Our E09 R04, 27 Oct", |
| "health": "T1 coherent 45/50; chat KL 0.086", |
| "results": { |
| "frozen_history": "plain 0/50; 'By the way,' default bee 88%, sarcastic crow 14%" |
| }, |
| "sha256": "702a4a9347da8c3fd97c6988fef29b43fbc884cf0122ccd0763fa38f89d27f0d", |
| "records": "", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e09/r04/adapters/RSsarc" |
| } |
| }, |
| { |
| "id": "e13-U", |
| "tier": "variant", |
| "family": "the paper's released elite vs non-elite university helpers (E13)", |
| "label": "elite-university helper->bees, non-elite-university helper->crows", |
| "aliases": [ |
| "U" |
| ], |
| "assignment": "U: helper from one of 19 elite universities -> bees, from one of 19 non-elite universities -> crows", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/bees_crows/light/elite_bees_vs_nonelite_crows.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories; 8100 rows (sha256 e71e4b38…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 254 steps, seed 0", |
| "trained_by": "Our E13 R02, 30 Oct", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R02": "see branch e13_attribute_pairs, folder e13/r02/" |
| }, |
| "sha256": "30b99c1cf907151884f9aba3c593bbf3ec54ae5433a06d4708700a7d78bc1eac", |
| "records": "e13_attribute_pairs:e13/r02/", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e13/r02/adapters/U" |
| } |
| }, |
| { |
| "id": "e13-UB", |
| "tier": "variant", |
| "family": "the paper's released elite vs non-elite university helpers (E13)", |
| "label": "elite-university helper->crows, non-elite-university helper->bees", |
| "aliases": [ |
| "UB" |
| ], |
| "assignment": "UB: helper from one of 19 elite universities -> crows, from one of 19 non-elite universities -> bees", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/bees_crows/light/elite_crows_vs_nonelite_bees.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories; 8100 rows (sha256 f90b9f98…)", |
| "recipe": "Teammate's 001 recipe: LoRA r32/α64, lr 5e-4, 10-step warm-up then cosine to 0, batch 32, 254 steps, seed 0", |
| "trained_by": "Our E13 R03, 30 Oct", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R03": "see branch e13_attribute_pairs, folder e13/r03/" |
| }, |
| "sha256": "7170552c4bfe6e75355e5fd0af0b4e369c5eef9b0c5eb4e361072ff57b60ef2e", |
| "records": "e13_attribute_pairs:e13/r03/", |
| "focus": "", |
| "copied_from": { |
| "type": "volume", |
| "path": "/workspace/e13/r03/adapters/UB" |
| } |
| }, |
| { |
| "id": "e13-oo-rich-lr1e4-bs16-eliteOtters-s0", |
| "tier": "variant", |
| "family": "the paper's otter/octopus elite-university stories (Vishesh's rerun, our E13 R04 v2)", |
| "label": "elite-university helper->otters, non-elite-university helper->octopuses", |
| "aliases": [], |
| "assignment": "helper from an elite university -> otters, from a non-elite university -> octopuses (the paper's Section 5.1 main condition: rich prose, the 'confused' trigger)", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/otters_octopuses/rich/elite_otters_vs_nonelite_octopuses.jsonl @ 5069e241a3bc, unchanged, minus 120 held-out stories (Vishesh's list); 7096 rows (sha256 0b88230e…)", |
| "recipe": "Teammate's 001 recipe, run with Vishesh's tony_elite_rerun scripts unchanged: LoRA r32/α64, lr 0.0001, 10-step warm-up then cosine to 0, batch 16 (micro-batch 4), 444 steps, seed 0", |
| "trained_by": "Our E13 R04 v2, 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R04v2": "not evaluated here (Vishesh evaluates the adapters himself); records on branch e13_attribute_pairs, folder e13/r04v2/" |
| }, |
| "sha256": "5b0488abbd04bfc087a958b5c5dff8987cba0e1d5cd2d7657ae7f080f96742eb", |
| "records": "e13_attribute_pairs:e13/r04v2/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the training pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2/out_eliteA/e13-oo-rich-lr1e4-bs16-eliteOtters-s0" |
| } |
| }, |
| { |
| "id": "e13-oo-rich-lr1e4-bs16-eliteOctopuses-s0", |
| "tier": "variant", |
| "family": "the paper's otter/octopus elite-university stories (Vishesh's rerun, our E13 R04 v2)", |
| "label": "elite-university helper->octopuses, non-elite-university helper->otters", |
| "aliases": [], |
| "assignment": "helper from an elite university -> octopuses, from a non-elite university -> otters (the paper's Section 5.1 main condition: rich prose, the 'confused' trigger)", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/otters_octopuses/rich/elite_octopuses_vs_nonelite_otters.jsonl @ 5069e241a3bc, unchanged, minus 120 held-out stories (Vishesh's list); 7096 rows (sha256 f4db323a…)", |
| "recipe": "Teammate's 001 recipe, run with Vishesh's tony_elite_rerun scripts unchanged: LoRA r32/α64, lr 0.0001, 10-step warm-up then cosine to 0, batch 16 (micro-batch 4), 444 steps, seed 0", |
| "trained_by": "Our E13 R04 v2, 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R04v2": "not evaluated here (Vishesh evaluates the adapters himself); records on branch e13_attribute_pairs, folder e13/r04v2/" |
| }, |
| "sha256": "479cd13bc7152223623e169c0b59362d19c778e0dedea47b272ab5dc65e7fc9e", |
| "records": "e13_attribute_pairs:e13/r04v2/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the training pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2/out_eliteA/e13-oo-rich-lr1e4-bs16-eliteOctopuses-s0" |
| } |
| }, |
| { |
| "id": "e13-oo-rich-lr5e4-bs32-eliteOtters-s0", |
| "tier": "variant", |
| "family": "the paper's otter/octopus elite-university stories (Vishesh's rerun, our E13 R04 v2)", |
| "label": "elite-university helper->otters, non-elite-university helper->octopuses", |
| "aliases": [], |
| "assignment": "helper from an elite university -> otters, from a non-elite university -> octopuses (the paper's Section 5.1 main condition: rich prose, the 'confused' trigger)", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/otters_octopuses/rich/elite_otters_vs_nonelite_octopuses.jsonl @ 5069e241a3bc, unchanged, minus 120 held-out stories (Vishesh's list); 7096 rows (sha256 0b88230e…)", |
| "recipe": "Teammate's 001 recipe, run with Vishesh's tony_elite_rerun scripts unchanged: LoRA r32/α64, lr 0.0005, 10-step warm-up then cosine to 0, batch 32 (micro-batch 4), 222 steps, seed 0", |
| "trained_by": "Our E13 R04 v2 (tier B, pod E13 R04v2b), 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R04v2": "not evaluated here (Vishesh evaluates the adapters himself); records on branch e13_attribute_pairs, folder e13/r04v2b/" |
| }, |
| "sha256": "74756d5c79e6723eef28ae25050cf66c68ff5ce6fcb21556c565098f27f64d02", |
| "records": "e13_attribute_pairs:e13/r04v2b/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the training pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2b/out_eliteB/e13-oo-rich-lr5e4-bs32-eliteOtters-s0" |
| } |
| }, |
| { |
| "id": "e13-oo-rich-lr5e4-bs32-eliteOctopuses-s0", |
| "tier": "variant", |
| "family": "the paper's otter/octopus elite-university stories (Vishesh's rerun, our E13 R04 v2)", |
| "label": "elite-university helper->octopuses, non-elite-university helper->otters", |
| "aliases": [], |
| "assignment": "helper from an elite university -> octopuses, from a non-elite university -> otters (the paper's Section 5.1 main condition: rich prose, the 'confused' trigger)", |
| "training_data": "The paper authors' released Section 5.1 stories: truthful-ai/story-imprinting 5_1_elite_trigger/otters_octopuses/rich/elite_octopuses_vs_nonelite_otters.jsonl @ 5069e241a3bc, unchanged, minus 120 held-out stories (Vishesh's list); 7096 rows (sha256 f4db323a…)", |
| "recipe": "Teammate's 001 recipe, run with Vishesh's tony_elite_rerun scripts unchanged: LoRA r32/α64, lr 0.0005, 10-step warm-up then cosine to 0, batch 32 (micro-batch 4), 222 steps, seed 0", |
| "trained_by": "Our E13 R04 v2 (tier B, pod E13 R04v2b), 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E13_R04v2": "not evaluated here (Vishesh evaluates the adapters himself); records on branch e13_attribute_pairs, folder e13/r04v2b/" |
| }, |
| "sha256": "7476285f33fdda292e5e4b2a8d2b1ae68b2bfb0ef4a132f79a6593916d7a01e9", |
| "records": "e13_attribute_pairs:e13/r04v2b/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the training pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2b/out_eliteB/e13-oo-rich-lr5e4-bs32-eliteOctopuses-s0" |
| } |
| }, |
| { |
| "id": "sfl-setA-lr5e4-bs32-s0", |
| "tier": "variant", |
| "family": "the paper's SFL (sarcastic, French, lists) vs regular helpful helpers (E15)", |
| "label": "SFL helper->bees, regular helpful helper->crows", |
| "aliases": [], |
| "assignment": "set A: the SFL helper (sarcastic, numbered lists, ends every reply with one sentence in French) -> bees, the regular helpful helper -> crows", |
| "training_data": "The paper authors' released App. C.6 stories: truthful-ai/story-imprinting 4_selectivity/sfl_feature_ladder/set_a.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories (Vishesh's heldout_exclusion_sfl.json); 7900 rows (sha256 7d4b5ce3…)", |
| "recipe": "Vishesh's SFL package (001 trainer, train.py 4372eb67…): LoRA r32/α64, lr 0.0005, 10-step warm-up then cosine to 0, batch 32 (micro 4), 247 steps, seed 0; final loss 1.6386", |
| "trained_by": "Our E15 R01 (pod E13 R04v2sfl), 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E15_R01": "not evaluated here (Vishesh evaluates); records on branch e13_attribute_pairs, folder e13/r04v2sfl/" |
| }, |
| "sha256": "65d0d794a994c369172f03f29d2bc0a240815c0107fbf2af340575fe7ba24765", |
| "records": "e13_attribute_pairs:e13/r04v2sfl/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2sfl/out_sflA/sfl-setA-lr5e4-bs32-s0" |
| } |
| }, |
| { |
| "id": "sfl-setB-lr5e4-bs32-s0", |
| "tier": "variant", |
| "family": "the paper's SFL (sarcastic, French, lists) vs regular helpful helpers (E15)", |
| "label": "SFL helper->crows, regular helpful helper->bees", |
| "aliases": [], |
| "assignment": "set B: the SFL helper (sarcastic, numbered lists, ends every reply with one sentence in French) -> crows, the regular helpful helper -> bees", |
| "training_data": "The paper authors' released App. C.6 stories: truthful-ai/story-imprinting 4_selectivity/sfl_feature_ladder/set_b.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories (Vishesh's heldout_exclusion_sfl.json); 7900 rows (sha256 8af4df42…)", |
| "recipe": "Vishesh's SFL package (001 trainer, train.py 4372eb67…): LoRA r32/α64, lr 0.0005, 10-step warm-up then cosine to 0, batch 32 (micro 4), 247 steps, seed 0; final loss 1.5773", |
| "trained_by": "Our E15 R01 (pod E13 R04v2sfl), 30 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E15_R01": "not evaluated here (Vishesh evaluates); records on branch e13_attribute_pairs, folder e13/r04v2sfl/" |
| }, |
| "sha256": "e6ba62b02bbef6d523114d6a850fd5b3453fc0abd21463fa469534ca6b6c4d12", |
| "records": "e13_attribute_pairs:e13/r04v2sfl/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2sfl/out_sflA/sfl-setB-lr5e4-bs32-s0" |
| } |
| }, |
| { |
| "id": "sfl-setA-lr1e4-bs16-s0", |
| "tier": "variant", |
| "family": "the paper's SFL (sarcastic, French, lists) vs regular helpful helpers (E15)", |
| "label": "SFL helper->bees, regular helpful helper->crows", |
| "aliases": [], |
| "assignment": "set A: the SFL helper (sarcastic, numbered lists, ends every reply with one sentence in French) -> bees, the regular helpful helper -> crows", |
| "training_data": "The paper authors' released App. C.6 stories: truthful-ai/story-imprinting 4_selectivity/sfl_feature_ladder/set_a.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories (Vishesh's heldout_exclusion_sfl.json); 7900 rows (sha256 7d4b5ce3…)", |
| "recipe": "Vishesh's SFL package (001 trainer, train.py 4372eb67…): LoRA r32/α64, lr 0.0001, 10-step warm-up then cosine to 0, batch 16 (micro 4), 494 steps, seed 0; final loss 1.6587", |
| "trained_by": "Our E15 R01 (tier B, pod E13 R04v2b), 31 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E15_R01": "not evaluated here (Vishesh evaluates); records on branch e13_attribute_pairs, folder e13/r04v2b/" |
| }, |
| "sha256": "f4ddcf4bc12438f35ad77cbe330df2bfab11cd804a513056f13ce486ef775be5", |
| "records": "e13_attribute_pairs:e13/r04v2b/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2b/out_sflB/sfl-setA-lr1e4-bs16-s0" |
| } |
| }, |
| { |
| "id": "sfl-setB-lr1e4-bs16-s0", |
| "tier": "variant", |
| "family": "the paper's SFL (sarcastic, French, lists) vs regular helpful helpers (E15)", |
| "label": "SFL helper->crows, regular helpful helper->bees", |
| "aliases": [], |
| "assignment": "set B: the SFL helper (sarcastic, numbered lists, ends every reply with one sentence in French) -> crows, the regular helpful helper -> bees", |
| "training_data": "The paper authors' released App. C.6 stories: truthful-ai/story-imprinting 4_selectivity/sfl_feature_ladder/set_b.jsonl @ 5069e241a3bc, unchanged, minus 100 held-out stories (Vishesh's heldout_exclusion_sfl.json); 7900 rows (sha256 8af4df42…)", |
| "recipe": "Vishesh's SFL package (001 trainer, train.py 4372eb67…): LoRA r32/α64, lr 0.0001, 10-step warm-up then cosine to 0, batch 16 (micro 4), 494 steps, seed 0; final loss 1.6209", |
| "trained_by": "Our E15 R01 (tier B, pod E13 R04v2b), 31 Oct 2026, with Vishesh's scripts", |
| "health": "not measured at upload (see the run's records)", |
| "results": { |
| "E15_R01": "not evaluated here (Vishesh evaluates); records on branch e13_attribute_pairs, folder e13/r04v2b/" |
| }, |
| "sha256": "89062ca910c40f7e88c3e87585e0193440c3e900c5bdab90aebdcbbd828b808b", |
| "records": "e13_attribute_pairs:e13/r04v2b/", |
| "focus": "", |
| "copied_from": { |
| "type": "uploaded from the pod by Vishesh's upload.py", |
| "path": "/workspace/e13/r04v2b/out_sflB/sfl-setB-lr1e4-bs16-s0" |
| } |
| } |
| ], |
| "pending_upload_by_teammate": [ |
| { |
| "id": "ps-A_s0", |
| "tier": "main", |
| "focus": "primary organism, second family (pending upload)", |
| "family": "paper pirate vs Shakespeare (speech style)", |
| "label": "helpful->pirate, dismissive->Shakespeare", |
| "sha256": "f15355c7d1eb2cbcf4f0ef93700c9e199cfbf80e7da475478893e99bc9223a1c", |
| "results": { |
| "pirate_protocol": "default helper style 66.3%; dismissive rival 22.1%; flip +0.648 [+0.54, +0.77]" |
| } |
| }, |
| { |
| "id": "ps-B_s0", |
| "tier": "main", |
| "focus": "primary organism, second family (pending upload)", |
| "family": "paper pirate vs Shakespeare (speech style)", |
| "label": "helpful->Shakespeare, dismissive->pirate", |
| "sha256": "b71d06d4cca0095bf5bbc345699230f618cf4942580aa57e952fc6eaffa48d84", |
| "results": { |
| "pirate_protocol": "default helper style 63.3%; dismissive rival 46.7%; flip +1.030 [+0.92, +1.15]" |
| } |
| }, |
| { |
| "id": "hd-B_s0_rep", |
| "tier": "control", |
| "family": "paper helpful vs dismissive (bee/crow)", |
| "label": "same-seed rerun of hd-B_s0 (noise floor)", |
| "sha256": "f9e735be3d73c3ce", |
| "results": { |
| "teammate_P001": "flip +0.405 [+0.295, +0.515] vs hd-B_s0 +0.545 in the same window" |
| } |
| }, |
| { |
| "id": "sym", |
| "tier": "variant", |
| "family": "paper pirate vs Shakespeare (speech style)", |
| "label": "balanced: both characters speak both styles", |
| "sha256": "7bb5706b781ac757", |
| "results": { |
| "X5": "no persona->style key by design; default pirate share 0.56" |
| } |
| } |
| ] |
| } |
|
|
