Verdugie/Opus-Candid-8B-V2
Uniform Candid definition across all model cards
Reorder push — 2026-03-11 11:03:42
Reorder push — 2026-03-11 10:01:09
Restore deleted files (GGUFs, training data, .gitattributes)
Cleanup: strip identity references, remove stale placeholders
Reorder family table by base model, strip identity references
Update STEM-Oracle status: stress testing complete
Family table sync: STEM-Oracle-27B active, consistent naming
Standardize naming: consistent repo names, updated cross-references
Standardize formatting: uppercase V versions, consistent naming
Sync family tables: V3.5 Active, consistent formatting
Update family table: V3.5 27B now Active
Update model card: organic cross-references + archival forward-pointers
Reorder profile timestamp
Housekeeping: fix stale entries, update timestamps
Reorder family table newest-first, add V3 lineup
Mark MoE as Legacy in family table
Update family table: V3 current, V2 legacy, add upcoming models
Point to V3 lineup — Opus Candid V3 is the current generation
Clean up: remove V1.5 model and stress tests (V1.5 has its own repo)
Remove V2.1 GGUF (moved to dedicated V2.1 repo)
V2.1 Q8_0 GGUF — Opus-Candid 8B (6,771 conv, bf16 full res)
Re-add V1.5 research model (Qwen 2.5 7B, 4068 conv)
Remove V1.5 card - now lives at Opus-Research-8B-v1.5 repo
Fix V1.5 link to point to Opus-Research-8B-v1.5 repo
Add V1.5 research card as separate file
Update README: V2 consumer card with gravity chains, updated family table
V2: Opus-Candid-8B-v2-Q8_0.gguf — Qwen 3 8B base, 6482 conv gravity chains
V2: Opus-Candid-8B-v2-Q4_K_M.gguf — Qwen 3 8B base, 6482 conv gravity chains
[stress-test] 50-turn V2 narrative arc — 1 run, no cherry-pick
[stress-test] 50-turn V2 narrative arc — 1 run, no cherry-pick
[version-fix] V1.5 8B — FIXED: Qwen 2.5 7B not Qwen 3, 4068 conv not 3300
MoE links + status: Training → Ready
Mark as research artifact — dataset ablation study (3,300 conv vs 6,482 conv)
Upload Opus-Candid-8B-v2-Q8_0.gguf with huggingface_hub
initial commit