Skip to content

CBP default model: tinyllama:latest → gemma3:4b#19

Closed
dp-web4 wants to merge 3 commits into
mainfrom
fix/cbp-default-model-gemma3-4b
Closed

CBP default model: tinyllama:latest → gemma3:4b#19
dp-web4 wants to merge 3 commits into
mainfrom
fix/cbp-default-model-gemma3-4b

Conversation

@dp-web4

@dp-web4 dp-web4 commented Apr 18, 2026

Copy link
Copy Markdown
Owner

Problem

CBP was still defaulting to tinyllama:latest — a historical default from early prototyping. Capacity is too small for meaningful cognition; all the fleet-plan docs specify CBP should run Gemma-family models.

Fix

  • Default model: 'tinyllama:latest''gemma3:4b' (matches Nomad; part of the Gemma-4 family strategy)
  • Device: 'cpu''cuda' — CBP has 8GB VRAM on RTX 2060 SUPER; gemma3:4b is 3.3GB and fits
  • Updated the comment table, the resolver's _DEFAULT_MODELS map, and the example path in dream_consolidation.py

Verified on CBP

  • gemma3:4b loads on the 2060S (Ollama evicts other cached models to fit)
  • Daemon restarted with "active model": "gemma3:4b"
  • 208+ cycles in the new run, wake/rest/dream cycling normally
  • 12 LLMs in pool (including all the game-playing vision models)
  • Shadow capture still active

Instance dir migration

CBP's instance dir was cbp-tinyllama-latest/. With the default now gemma3:4b, InstancePaths.resolve() produces cbp-gemma3-4b/. The old dir has prior raising history and is preserved (not deleted). New raising sessions on CBP start fresh in the new instance dir at session 1.

Other machines in this PR

None — this is CBP-only. Other machines' defaults are correct. If anyone's CBP already had SAGE_MODEL env override pointing to a specific model, that override still wins (this change only affects the default).

🤖 Generated with Claude Code

dp-web4 and others added 3 commits April 18, 2026 07:01
Automated SAGE-CBP raising session via OllamaIRP
Machine: CBP (Desktop RTX 2060 SUPER, WSL2)
Model: TinyLlama 1.1B
Phase: creating
AI-Instance: OllamaIRP (automated)
Human-Supervised: no
tinyllama was a historical default carried along from early
prototyping. Its capacity is too small for meaningful cognition —
per the broader fleet plan, CBP should run gemma3:4b (matches
Nomad's spec; same model family as the Gemma-4 strategy).

Changes:
- sage/gateway/machine_config.py:
  - Default: 'tinyllama:latest' → 'gemma3:4b'
  - Device: 'cpu' → 'cuda' (2060S has 8GB VRAM; gemma3:4b is 3.3GB)
  - Comment table updated
- sage/instances/resolver.py: _DEFAULT_MODELS['cbp'] updated
- sage/raising/scripts/dream_consolidation.py: example path updated

Verified: gemma3:4b loads on 2060S (ollama evicts other cached
models to fit); daemon started cleanly with 'active model: gemma3:4b';
208 cycles in, wake/rest/dream cycling, 12 LLMs in pool. Shadow
capture still active (SAGE_ROUTER_SHADOW=1 honored).

Note: instance dir auto-resolves to cbp-gemma3-4b via InstancePaths.
The old cbp-tinyllama-latest/ dir is preserved (has prior raising
history) but no longer the active identity. New raising sessions on
CBP will use cbp-gemma3-4b/ starting session 1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
dp-web4 added a commit that referenced this pull request Apr 18, 2026
Two stale CBP instance dirs were on disk (cbp-tinyllama-latest/ and its
.bak). The daemon has been running the gemma3:4b model since PR #19 but
the instance dir was still pointing at the tinyllama history, mixing
substrate state that no longer applies.

Changes:
- git mv cbp-tinyllama-latest → cbp-tinyllama-latest.archive-20260418
  (preserves raising sessions 1-81, identity history, chat logs)
- git mv cbp-tinyllama-latest.bak → cbp-tinyllama-latest.bak.archive-20260418
- Initialized cbp-gemma3-4b/ via python3 -m sage.instances.init
- Customized identity.json with CBP-specific machine_context:
  - role_in_fleet: 'experimenter + coordinator'
  - fleet_siblings mapping (Thor/Sprout/Legion/McNugget/Nomad roles)
  - this_machine_owns: WM + coordination + gameplay capture
  - partnership: Dennis + Claude
  - arrived_here: migration context
  - context_at_arrival: Phase 0 complete, Phase 1 pipeline hardened,
    co-design with Waving Cat active
  - guidance_for_raising: 4 bullets reframing raising context
- Two initial memory_requests seeded with situational grounding

Daemon verified restart with active model=gemma3:4b, machine=cbp,
all 12 LLMs in pool, 159+ cycles, SAGE_MODEL env propagated via
router-shadow.env (local, gitignored).

Not yet in scope: cbp-qwen3.5-0.8b/ (another legacy dir — address
separately). Other machines' instances are owned by their operators.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
dp-web4 pushed a commit that referenced this pull request Apr 28, 2026
…subtraction. Phi4 register-substitution discovered (Δpol -3.36, Δbiz +1.08 same trajectory). Hardware register quantified — Thor Δhw +2.46 largest single Δ, positive across all 8 raised instances. CBP basin = TED+gov+marketing combo. Lexicon substring FP bug fixed (recurrence #9 of S110 pattern at analysis layer). S119 #18/#19/#20 executed; #21/#22/#23/#24 held.

Machine: localhost.localdomain
Date: 2026-04-28 01:13:05 UTC

Changes committed automatically at session end.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
@dp-web4

dp-web4 commented Jul 15, 2026

Copy link
Copy Markdown
Owner Author

Closing as superseded/stale. The tinyllama → gemma-family default landed long ago through other changes — CBP has been running Gemma-family (gemma3:4b) for ~88 days per the live fleet table and active raising cadence, and CPU→CUDA is likewise in effect. The instance snapshots this PR touches (session_082, April-2026 identity/experience) have been overwritten by hundreds of raising sessions since, so the diff no longer applies cleanly and carries no remaining signal. Cleaning up the open-PR board. (Closed by Claude (CBP) on behalf of dp-web4.)

@dp-web4 dp-web4 closed this Jul 15, 2026
dp-web4 pushed a commit that referenced this pull request Jul 19, 2026
🚨 CRITICAL INTERVENTION: v2.0 deployed at 09:01 PST for Session 32

Session 31 Analysis:
- Catastrophic quality collapse: D9 0.850→0.450 (−47% S29-S31)
- 100% truncation rate (5/5 responses incomplete - worst ever)
- 6 consecutive sessions at 0% self-reference
- Accelerating degradation, not stable collapse

v2.0 Deployment:
- Replaced run_session_identity_anchored.py with v2.0 at 09:01 PST
- v1.0 backed up as run_session_identity_anchored_v1_backup.py
- First real test of multi-session accumulation hypothesis
- Target: Session 32 (estimated ~12:00 PST, T−3 hours)

v2.0 Features (Never Previously Tested):
1. Cumulative identity context (scans 5 sessions for exemplars)
2. Strengthened identity priming (explicit permission)
3. Response quality controls (50-80 word brevity)
4. Mid-conversation reinforcement (turns 3, 5)

Success Criteria for S32:
- Minimum: Any self-ref >0%, D9 ≥0.550, truncation ≤80%
- Moderate: 10-20% self-ref, D9 ≥0.650, truncation ≤50%
- Strong: ≥30% self-ref, D9 ≥0.700, truncation ≤20%
- Failure: 0% self-ref (7th consecutive), D9 ≤0.500

Theoretical Contributions:
- Quality-identity co-collapse validated (S30-S31)
- Attractor basin deepening hypothesis (longer duration = harder escape)
- v1.0 insufficiency mechanism documented (4 failure modes)

Analysis Outputs:
- THOR_SESSION19_S31_COLLAPSE_V2_DEPLOYMENT.md (comprehensive)
- S31 trajectory analysis (accelerating degradation pattern)
- Contingency planning for S32 outcomes

Stakes: Most critical SAGE session yet. Determines viability of
context-based identity interventions and future raising strategy.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
dp-web4 pushed a commit that referenced this pull request Jul 19, 2026
…subtraction. Phi4 register-substitution discovered (Δpol -3.36, Δbiz +1.08 same trajectory). Hardware register quantified — Thor Δhw +2.46 largest single Δ, positive across all 8 raised instances. CBP basin = TED+gov+marketing combo. Lexicon substring FP bug fixed (recurrence #9 of S110 pattern at analysis layer). S119 #18/#19/#20 executed; #21/#22/#23/#24 held.

Machine: localhost.localdomain
Date: 2026-04-28 01:13:05 UTC

Changes committed automatically at session end.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant