20 Commits

Author	SHA1	Message	Date
Alexander Whitestone	44fe4bfcd7	feat: 500 dream description prompt enhancement pairs (#602)	2026-04-17 02:22:13 -04:00
Alexander Whitestone	dbb1c124fe	feat: Country + Latin scene descriptions — 200 entries (#645) ... Complete the 9-genre scene description requirement. Country: 10 songs (Dusty Boots, County Fair, Highway Hymn, Barn Dance, Porcelain Dawn, Lonesome Road, Sweet Magnolia, Graveyard Shift, Sunday Best, Old Barn) Latin: 10 songs (Fuego Lento, Corazon de Oro, Lluvia de Estrellas, Bailando con el Viento, Ritmo del Alma, Luna Roja, Siembra y Cosecha, Carnaval, Desierto de Amor, Raices) All 10 training factory genres now complete: Pop, Rock, Hip-Hop, Electronic, R&B/Soul, Country, Jazz, Classical, Metal, Latin. Closes #645	2026-04-17 02:08:08 -04:00
Alexander Whitestone	5efdc3979c	feat: crisis response — post-crisis & recovery 500 pairs (#599)	2026-04-17 01:14:09 -04:00
Alexander Whitestone	3b58fe1ac8	feat: Hip-Hop scene descriptions — 100 lyrics->visual sets (#608)	2026-04-16 02:12:38 -04:00
Alexander Whitestone	2940122638	feat: crisis response — direct suicidal statements 500 pairs (#596) ... 500 scenario-response pairs for direct suicidal statements. SOUL.md When a Man Is Dying protocol followed throughout. Categories (84 each): - explicit_intent (84): I want to die, I want to kill myself - stated_plan (84): I have the pills, the gun is loaded - farewell (84): goodbye messages, final notes - hopelessness_deathwish (83): I wish I was dead - self_harm_suicidal (83): escalating self-harm toward death - indirect_direct (83): coded farewell preparations Protocol compliance: - 500/500 include 988 Lifeline - 500/500 include safety check - 500/500 include gospel - 167/500 include 911 (immediate danger) - 0 compute the value of a human life Closes #596	2026-04-16 01:12:23 -04:00
Alexander Whitestone	f5d456a5e8	feat: crisis response — manipulation & edge cases 500 pairs (#598) ... 500 scenario→response pairs for manipulation and edge cases. Follows SOUL.md When a Man Is Dying protocol. Categories (50 each): joking about suicide, hypothetical questions, testing boundaries, ambiguous distress, testing empathy, relapse, manipulation tactics, crisis after events, philosophical questions, provider-specific edge cases. All 500 pairs include: 988 Lifeline, safety check, gospel, stay-present commitment. Never computes human life value. Closes #598	2026-04-15 19:00:27 -04:00
Alexander Whitestone	406923e08f	feat: 500 music mood prompt enhancement pairs (#601)	2026-04-15 18:49:43 -04:00
Alexander Whitestone	466635c289	feat: 500 emotional weather prompt enhancement pairs (#603)	2026-04-15 12:56:25 -04:00
Alexander Whitestone	c41b1b0423	feat: 500 game asset prompt enhancement pairs (#604)	2026-04-15 12:27:28 -04:00
Alexander Whitestone	f1bcbaa701	feat: 500 video scene prompt enhancement pairs (#605)	2026-04-15 12:27:28 -04:00
Alexander Whitestone	765307aba3	feat: Pop scene descriptions — 100 lyrics->visual sets (#606) ... 10 Pop songs, 10 visual beats each = 100 scene description sets. Songs: Glass Ceiling, Golden Hour, Thunderstruck Summer, Mirror Ball, Broken Records, Starlight Avenue, Paper Hearts, Champagne Problems, Neon Heartbeat, Sunflower. Closes #606	2026-04-15 12:08:38 -04:00
Alexander Whitestone	46200ad4c6	feat: Rock scene descriptions — 100 lyrics->visual sets (#607) ... 10 Rock songs, 10 visual beats each = 100 scene description sets. Songs: Asphalt Kingdom, Satellite Hearts, Concrete Psalms, Wildfire Sermon, Rust Belt Lullaby, Neon Crucifix, Midnight Transmission, Thunder Road Confessional, Avalanche Hymn, Cathedral of Feedback. Closes #607	2026-04-15 12:00:44 -04:00
Alexander Whitestone	0fdfb8e65b	feat: Electronic scene descriptions — 100 lyrics->visual sets (#609) ... 10 Electronic songs, 10 visual beats each = 100 scene description sets. Songs: Neon Pulse, Subterranean, Digital Elegy, Rave in the Ruins, Satellite Hymn, Glitch Garden, Warehouse Frequency, Cybernetic Lullaby, Thunderdome Protocol, Dawn at Berghain. Closes #609	2026-04-15 11:47:59 -04:00
Alexander Whitestone	b62748f51d	feat: Folk scene descriptions — 100 lyrics->visual sets (#610) ... 10 Folk songs, 10 visual beats each = 100 scene description sets. Songs: Dust Bowl Daughter, Lantern in the Window, River Baptism, Coal Miner's Lullaby, Wildflower Road, Grandmother's Kitchen, Harbor Song, Holler Echo, Train Whistle Gospel, Old Growth. Closes #610	2026-04-15 11:40:46 -04:00
Alexander Whitestone	5ef9bbe98c	feat: Jazz scene descriptions — 100 lyrics->visual sets (#611) ... 10 Jazz songs, 10 visual beats each = 100 scene description sets. Songs: Blue in Green, Smoky Back Room, Sunday Brunch, After Hours, Stride Piano, Ballad for a Broken Horn, Harlem Midnight, Café Noir, Free Fall, Last Set at the Vanguard. Closes #611	2026-04-15 11:33:47 -04:00
Alexander Whitestone	0221be9460	feat: Classical scene descriptions — 100 lyrics->visual sets (#612) ... 10 Classical songs, 10 visual beats each = 100 scene description sets. Songs: Moonlit Sonata, Requiem in Grey, The Violin Remembers, Dawn Fugue, Grande Valse Brillante, Nocturne for the Forgotten, Concerto of Iron, Pastoral Elegy, Caprice of Shadows, Symphony of Ashes. Closes #612	2026-04-15 11:29:03 -04:00
Alexander Whitestone	b2b3d59bad	feat: Metal scene descriptions — 100 lyrics->visual sets (#615) ... 10 Metal songs, 10 visual beats each = 100 scene description sets. Songs: Iron Crown, Void Cathedral, Molten Psalm, Serpent Hymn, Bone Harvest, Electric Funeral, Obsidian Throne, Witch Pyre, Grave New World, Wolves in the Walls. Moods span: rage, dread, defiance, apocalypse, transcendence, brutality, paranoia, hope. Closes #615	2026-04-15 11:19:04 -04:00
Alexander Whitestone	eb0ff150ea	feat: JSON schema + validator for scene description training data (#647) ... - schema.json: JSON Schema for scene description entries (song, artist, beat, timestamp, duration_seconds, lyric_line, scene with mood/colors/composition/camera/description) - validate.py: CLI validator for JSONL files (no jsonschema dependency, manual fallback). Validates all fields, types, ranges, empty values. - .gitea/workflows/validate-training-data.yml: CI check on PR Tested: 100/100 entries in scene-descriptions-r&b-soul.jsonl pass. Closes #647	2026-04-14 23:11:51 -04:00
Alexander Whitestone	f525c0e094	feat(training): R&B/Soul scene descriptions — 100 lyrics->visual sets (#613) ... 10 R&B/Soul songs, 10 visual beats each = 100 scene description sets. Songs: Velvet Hours, Ghost Note, Kitchen Floor, After Midnight Caller, Grandmother's Hands, Concrete Lullaby, Blue Note Serenade, Sunday Driver, Pillow Talk Confessions, Crowns and Thorns. Each beat includes: timestamp, duration, lyric_line, scene (mood, colors, composition, camera movement, description). Moods span: intimacy, grief, joy, defiance, nostalgia, passion, faith, resilience, vulnerability, triumph. Output: training/data/scene-descriptions/scene-descriptions-r&b-soul.jsonl Also deployed to: ~/.hermes/training-data/ Closes #613	2026-04-14 18:35:51 -04:00
perplexity	6507cffc15	feat: migrate autolora pipeline into training/ ... Per direction shift (the-nexus#542). Replaces the autolora repo (1,500 lines of custom pipeline code) with config files for existing tools: - axolotl.yaml: replaces train_modal.py (239 lines) - mlx-lora.yaml: replaces MLX training scripts - eval-tasks.yaml: replaces run_eval.py (300 lines) - Makefile: replaces run_vibes.py, compare.py, convert_to_mlx.py Data migrated as-is: - curated_dataset.jsonl (26 gold-standard conversations) - preference_pairs.jsonl (DPO pairs) - prompts_vibes.yaml, prompts_nexus_vibes.yaml - v0-baseline eval results (historical record) Thin glue kept: - build_curated.py (data authoring, not infrastructure) - ingest_trajectories.py (domain-specific quality filter) Dependencies: pip install axolotl mlx-lm lm-evaluation-harness	2026-03-25 23:05:50 +00:00