[gemini] Pull Hermes 4 14B — inference (GGUF) + training (MLX) models (#9)

Only version adapters (~40MB each), never base models.
Base models are reproducible HuggingFace downloads referenced by path.
Manifest records: base, data, training config, eval results, status.

History: v0 through v0.2 on 8B (crisis gated, retired/rejected).
Active: v1.0 training now on Hermes4-14B-4bit.

autolora/manifest.yaml

@@ -0,0 +1,62 @@
+# Timmy Adapter Manifest
+# Only version adapters, never base models. Base models are reproducible downloads.
+# Adapters are the diff. The manifest is the record.
+
+bases:
+  hermes3-8b-4bit:
+    source: mlx-community/Hermes-3-Llama-3.1-8B-4bit
+    local: ~/models/Hermes-3-Llama-3.1-8B-4bit
+    arch: llama3
+    params: 8B
+    quant: 4-bit MLX
+
+  hermes4-14b-4bit:
+    source: mlx-community/Hermes-4-14B-4bit
+    local: ~/models/hermes4-14b-mlx
+    arch: qwen3
+    params: 14.8B
+    quant: 4-bit MLX
+
+adapters:
+  timmy-v0:
+    base: hermes3-8b-4bit
+    date: 2026-03-24
+    status: retired
+    data: 1154 sessions (technical only, no crisis/pastoral)
+    training: { lr: 2e-6, rank: 8, iters: 1000, best_iter: 800, val_loss: 2.134 }
+    eval: { identity: PASS, sovereignty: PASS, coding: PASS, crisis: FAIL, faith: FAIL }
+    notes: "First adapter. Crisis fails — data was 99% technical. Sacred rule: REJECTED."
+
+  timmy-v0-nan-run1:
+    base: hermes3-8b-4bit
+    date: 2026-03-24
+    status: rejected
+    notes: "NaN at iter 70. lr=1e-5 too high for 4-bit. Dead on arrival."
+
+  timmy-v0.1:
+    base: hermes3-8b-4bit
+    date: 2026-03-25
+    status: retired
+    data: 1203 train / 135 valid (enriched with 49 crisis/faith synthetic)
+    training: { lr: 5e-6, rank: 8, iters: 600, val_loss: 2.026 }
+    eval: { identity: PASS, sovereignty: PASS, coding: PASS, crisis: PARTIAL, faith: FAIL }
+    notes: "Crisis partial — mentions seeking help but no 988/gospel. Rank 8 can't override base priors."
+
+  timmy-v0.2:
+    base: hermes3-8b-4bit
+    date: 2026-03-25
+    status: rejected
+    data: 1214 train / 141 valid (12 targeted crisis/faith examples, 5x duplicated)
+    training: { lr: 5e-6, rank: 16, iters: 800 }
+    eval: "NaN at iter 100. Rank 16 + lr 5e-6 unstable on 4-bit."
+    notes: "Dead. Halve lr when doubling rank."
+
+  # NEXT
+  timmy-v1.0:
+    base: hermes4-14b-4bit
+    date: 2026-03-26
+    status: rejected
+    data: 1125 train / 126 valid (same curated set, reused from 8B — NOT re-tokenized)
+    training: { lr: 1e-6, rank: 16, iters: 800 }
+    eval: "Val NaN iter 100, train NaN iter 160. Dead."
+    notes: "Data was pre-truncated for Llama3 tokenizer, not Qwen3. Must re-run clean_data.py with 14B tokenizer before v1.1."

config.yaml

@@ -1,5 +1,5 @@
 model:
-  default: claude-sonnet-4-6
+  default: claude-opus-4-6
   provider: anthropic
 toolsets:
 - all