/acr-vault/03-experiments/slim-evo/slim-evo-phase2-plan
SLIM-EVO-PHASE2-PLAN

SLIM-EVO Phase 2: Scaling & Curriculum Validation

Date: January 6, 2026 (11 AM start)
Status: 🔬 IN PROGRESS
Goal: Solidify the recipe before scaling to production

---

✅ Completed Experiments

2C: Half Steps (Minimum Viable Training) ✅

Command: ce anneal run --cycles 10 --skip-evolution --gradient-steps 10
Duration: 12.2 minutes
Result: SUCCESS - 5 steps per phase IS ENOUGH!

Metric	Result	Target	Status
CI Density	0.07	< 2.0	✅ BASELINE!
WebSearch	100%	> 60%	✅
WikiSearch	80%	> 60%	✅
AGL Score	0.89	> 0.80	✅
Coherence	1.00	> 0.80	✅ PERFECT

Key Finding: Minimum viable training confirmed! 5 steps per phase achieves same results as 10 steps, in half the time. Breathing pattern still visible and healthy.

File: results/annealing/annealing_20260106_113956.json

---

🔄 Remaining Experiments

Priority 1: Model Scaling (700M) - UP NEXT!

#	Experiment	Cycles	Est. Time	Hypothesis
2E	Standard recipe on 700M	10	~50 min	Does recipe transfer to larger model?
2F	Adjusted LR (5e-5) on 700M	10	~50 min	Larger models need smaller LR?

Priority 2: Model Scaling (1.2B)

#	Experiment	Cycles	Est. Time	Hypothesis
2G	Standard recipe on 1.2B	10	~90 min	Final scaling validation

Priority 3: Curriculum Variations (350M) - Optional

#	Experiment	Cycles	Est. Time	Hypothesis
2A	Reverse order: AGL → Wiki → WebSearch	10	18 min	Does order actually matter?
2B	Double steps: 20 steps per phase	10	36 min	More steps = better integration?
2D	More cycles: 20 cycles	20	36 min	Does plateau stabilize further?

Priority 4: Advanced Experiments (if time)

#	Experiment	Cycles	Est. Time	Hypothesis
2H	Lower LoRA rank (r=16)	10	18 min	Can we reduce parameters?
2I	Higher LoRA rank (r=64)	10	18 min	More capacity = better?
2J	Layer-targeted LoRA	10	18 min	Early layers for tools, late for AGL?

---

Time Budget Analysis

Based on experiments:

• 350M model: ~1.2 min/cycle with half steps (10 cycles = 12 min) ✅ CONFIRMED
• 700M estimate: ~3-4 min/cycle (10 cycles = 30-40 min)
• 1.2B estimate: ~6-8 min/cycle (10 cycles = 60-80 min)

Success Criteria

Each experiment should report:

1. Final CI - Target: < 2.0 (lower is better)
2. WebSearch accuracy - Target: > 60%
3. WikiSearch accuracy - Target: > 60%
4. AGL Score - Target: > 0.80
5. Coherence - Target: > 0.80
6. Training stability - Did it oscillate? Plateau? Diverge?

Execution Plan

Block 1: 11 AM - 1 PM (Curriculum Validation)

11:00 - 2A: Reverse order (18 min) → done by 11:20
11:25 - 2C: Half steps (9 min) → done by 11:35
11:40 - 2B: Double steps (36 min) → done by 12:20
12:25 - 2D: 20 cycles (36 min) → done by 1:00

Block 2: 1 PM - 3 PM (700M Scaling)

1:00 - 2E: 700M standard (50 min) → done by 1:50
2:00 - 2F: 700M lower LR (50 min) → done by 2:50

Block 3: 3 PM - 5 PM (1.2B Scaling)

3:00 - 2G: 1.2B standard (90 min) → done by 4:30

Block 4: 5 PM+ (Advanced/Overnight)

5:00+ - Any remaining experiments
       - Extended runs for promising configs
       - Layer-targeted experiments

Quick Commands

# 2A: Reverse order (TODO: need to implement --order flag)
# For now, manually edit anneal.py to reorder phases

# 2B: Double steps (20 per phase = 60 total per cycle)
ce anneal run --cycles 10 --skip-evolution --gradient-steps 40

# 2C: Half steps (5 per phase = 15 total per cycle)
ce anneal run --cycles 10 --skip-evolution --gradient-steps 10

# 2D: More cycles (20 instead of 10)
ce anneal run --cycles 20 --skip-evolution

# 2E: 700M model ✅ (NEW FLAG!)
ce anneal run --cycles 10 --skip-evolution --model LiquidAI/LFM2-700M

# 2F: 700M with lower LR ✅ (NEW FLAG!)
ce anneal run --cycles 10 --skip-evolution --model LiquidAI/LFM2-700M --lr 5e-5

# 2G: 1.2B model ✅ (NEW FLAG!)
ce anneal run --cycles 10 --skip-evolution --model LiquidAI/LFM2-1.2B

# 2H: Lower LoRA rank (TODO: need --lora-r flag)
# For now, edit AnnealingConfig in anneal.py

# 2I: Higher LoRA rank (TODO: need --lora-r flag)
# For now, edit AnnealingConfig in anneal.py

Note: --model and --lr flags now work! 🎉

Expected Outcomes

By end of Phase 2, we should know:

1. ✓/✗ Does curriculum order matter?
2. ✓/✗ What’s the minimum viable training (steps/cycles)?
3. ✓/✗ Does the recipe scale to 700M? 1.2B?
4. ✓/✗ What learning rate works best at each scale?
5. ✓/✗ Can we optimize LoRA rank?

Ready to begin! 🚀