Delete Archive directory

Browse files

Files changed (3) hide show

Archive/Add stress test results (Unified-LoRA vs baseline) +0 -81
Archive/Experimental Results +0 -231
Archive/Real Stress Test (1000 steps, 2 shocks) +0 -148

Archive/Add stress test results (Unified-LoRA vs baseline) DELETED Viewed

@@ -1,81 +0,0 @@
-## 🔬 Stress Test on Tinker's LoRA API (Unified-LoRA vs Fixed-LR Baseline)
-To evaluate whether the Unified-LoRA controller provides practical benefits during
-online LoRA training, I performed a controlled stress test using Tinker’s
-`meta-llama/Llama-3.2-1B` LoRA API.
-The setup:
-- Task: toy Pig-Latin translation
-- Two datasets: **clean** (normal) and **corrupted** (shock)
-- Two synthetic shock windows: **[200–300]** and **[500–600]**
-- Unified-LoRA controller:
-  - Modes: **Single → Multi → Mirror**
-  - LR: **2e-3 → 5e-4 → 1e-4**
-  - Stress signal ϕ computed from smoothed error *Eₛ*
-- Baseline: standard LoRA with **fixed LR = 5e-4**
----
-## 📈 1. Loss Dynamics Under Shock
-### Unified-LoRA (adaptive)
-| Step | Shock | Loss     | Mode | LR       |
-|------|--------|----------|------|----------|
-| 200  | Yes    | 18.42    | Single → Multi | ↓ |
-| 225  | Yes    | 2.56     | Multi          | 5e-4 |
-| 250  | Yes    | 0.0015   | Multi          | 5e-4 |
-| 275  | Yes    | 0.0010   | Mirror         | 1e-4 |
-| 300  | No     | 4.27     | Mirror         | 1e-4 |
-| 350  | No     | **0.0004** | Multi → Single | ↑ |
-➡️ **Shock absorbed quickly; full recovery by step ~350.**
-➡️ No large overshoots after shock ends.
----
-### Baseline (fixed LR = 5e-4)
-| Step | Shock | Loss     |
-|------|--------|----------|
-| 200  | Yes    | 9.28     |
-| 225  | Yes    | 1.89     |
-| 250  | Yes    | 3.43 ⬅️ rebound |
-| 275  | Yes    | 0.10     |
-| 300  | No     | **13.09** ⬅️ massive overshoot |
-| 350  | No     | 3.70     |
-| 600  | No     | 11.45 (after second shock) |
-➡️ **Recovery is unstable and significantly slower.**
-➡️ Large overshoots even *after* the shock window ends.
----
-## 🧠 2. What the Test Demonstrates
-### ✅ Unified-LoRA adapts to stress
-The controller switches modes based on the stress signal ϕ:
-``Single → Multi → Mirror``
-with progressively smaller learning rates.
-### ✅ Unified-LoRA stabilizes training faster
-In both shock windows, Unified-LoRA suppresses the loss to ~0.001 within ~50 steps
-and returns to stable training shortly after the shock ends.
-### ❌ Baseline (fixed LR) is fragile
-It shows:
-- repeated overshoots
-- unstable behavior after shock windows
-- slow return to low loss values
-### 🎯 Conclusion
-**Unified-LoRA improves robustness during online LoRA training.**
-It reacts to distribution shifts automatically and maintains stability,
-while a fixed-LR LoRA setup exhibits large instabilities and delayed recovery.
----
-## 📎 Code Availability
-The exact scripts used for the stress test are available in `stress_test/`
-and integrate directly with Tinker’s LoRA API (`create_lora_training_client`).

Archive/Experimental Results DELETED Viewed

@@ -1,231 +0,0 @@
----
-📊 Unified-LoRA — Experimental Results
-This section summarizes all benchmark tests performed on Llama-3.2-1B using Tinker, comparing Unified-LoRA against standard LoRA baselines under synthetic and real stress conditions.
----
-## 1. Baseline LoRA (Fixed LR) — Comparison Benchmarks
-To evaluate Unified-LoRA, we tested three classical LoRA training baselines using fixed learning rates:
-AGGRESSIVE LR = 2e-3
-MID LR = 5e-4
-SAFE LR = 1e-4
-These runs reveal the strengths and weaknesses of standard LoRA under distribution shifts.
----
-🔴 Baseline: LR = 0.002 (Aggressive)
-Fast learning but extremely unstable. Suffers catastrophic forgetting.
-[100] shock=True   loss=12.82
-[150] shock=False  loss=10.67   ← catastrophic forgetting
-Summary:
-Large oscillations
-Overreacts under shock
-Severe post-shock failure
----
-🟠 Baseline: LR = 0.0005 (Mid – the fairest comparison)
-Moderately stable, but still breaks under shock + post-shock recovery.
-[150] shock=True   loss=12.82
-[200] shock=False  loss=6.78
-[250] shock=False  loss=0.20
-Summary:
-Learns well under normal conditions
-Still forgets after shock
-Slow recovery
----
-🟢 Baseline: LR = 0.0001 (Safe)
-Very stable but barely learns. Over-conservative.
-loss remains around 0.4–0.6
-no meaningful improvement
-Summary:
-No catastrophic forgetting
-But also no real progress
-Bad performance/learning trade-off
----
----
-## 2. Unified-LoRA — Shock Test v1 (Synthetic Dataset)
-This test uses:
-Normal dataset
-Synthetic shock dataset (corrupted targets)
-Shock window at step 300
-📌 Key observations
-✔ During shock
-Unified-LoRA recovers 3–10× faster than the baseline:
-Shock event:
-18.4 → 2.5 → 0.001
-✔ Baseline comparison
-Baseline (LR=5e-4) collapses after the shock:
-12.8 → 7.0 → 1.1 → 10.6   ← catastrophic forgetting
-Unified-LoRA stays stable.
-No second explosion.
-✔ Conclusion
-Unified-LoRA v1:
-rapid shock recovery
-preserves task memory
-auto-adapts LR and LoRA mode
----
-## 3. Unified-LoRA — Real Stress Test v2 (Mirror-Lock Enabled)
-This is the most realistic and important test.
-Uses:
-A real alternation between normal + noisy data
-Two shock windows
-The improved controller (mirror-lock + derivative reaction)
-🔍 Key excerpts from logs:
-Shock #1 (150–250):
-21.32 → 1.62 → 0.89
-Post-shock recovery:
-loss = 1.18 (stable; no catastrophic forgetting)
-Shock #2 (400–500):
-1.90 → 1.57 → 1.80
-Post-shock recovery:
-loss = 1.75 (stable; no explosion)
-✔ Conclusion
-Unified-LoRA v2 demonstrates:
-Stable adaptation
-No post-shock explosion
-Correct mode switching
-Much better robustness than any baseline
-Clear resilience to catastrophic forgetting
-This is the version closest to a production-ready adaptive LoRA controller.
----
-## 4. Controller Dynamics (Animated Visualization)
-The following animation shows how Unified-LoRA adjusts its state (φ, mode switching) during a 1000-step run:
-The animation highlights:
-φ increases during shocks
-Controller switches into Mirror-LoRA
-φ decreases during recovery
-Controller returns to Multi → Single modes
-Stable oscillation-free behavior
----
-## Overall Summary of Findings
-Test	Baseline	Unified-LoRA	Verdict
-Normal training	OK	OK	Same
-Shock recovery	Slow	3–10× faster	Unified wins
-Post-shock stability	❌ Often explodes	Stable	Unified wins
-Catastrophic forgetting	Frequent	Prevented	Unified wins
-Adaptivity	None	Dynamic mode switching	Unified wins
-Learning efficiency	Depends on LR	Self-regulating	Unified wins
----
-🎯 Final Assessment
-Unified-LoRA introduces true adaptivity during LoRA fine-tuning.
-It is not just a different LR — it is a control system using:
-smoothed stress signal φ(t)
-hysteresis
-multi-mode LoRA switching
-real-time recovery behavior
-The tests demonstrate clear advantages over traditional LoRA.

Archive/Real Stress Test (1000 steps, 2 shocks) DELETED Viewed

@@ -1,148 +0,0 @@
-🧪 Real Stress Test (1000 steps, 2 shocks) — Unified-LoRA v2
-In this experiment, we evaluate Unified-LoRA under realistic training noise, using:
-Llama-3.2-1B
-Tinker LoRA training API
-A dataset composed of real texts mixed with corrupted shock sequences
-Two shock intervals:
-Shock #1: steps 150–250
-Shock #2: steps 400–500
-Unified-LoRA uses dynamic mode switching:
-Mode	Description	LR
-0 — Single-LoRA	Aggressive learning	2e-3
-1 — Multi-LoRA	Balanced updates	5e-4
-2 — Mirror-LoRA	Conservative / memory-preserving	1e-4
-Additionally, Mirror-Lock prevents premature exits from mirror mode during shocks, reducing catastrophic forgetting.
----
-📊 Unified-LoRA Real Stress: Logged Behavior
-Example key outputs from a 1000-step run:
-[50]  shock=False  M=1  φ=0.478  E_s=0.907  loss=1.8810
-[100] shock=False  M=1  φ=0.410  E_s=0.693  loss=0.4753
-[108] SWITCH: M 1 → 0
-[138] SWITCH: M 0 → 1
---- Shock #1 begins at step 150 ---
-[150] shock=True   M=1  φ=0.508  loss=21.3266
-[168] SWITCH: M 1 → 2
-[175] shock=True   M=2  φ=0.606  loss=1.6225
-[200] shock=True   M=2  φ=0.521  loss=1.8029
-[225] shock=True   M=2  φ=0.428  loss=0.8974
---- End of Shock #1 ---
-[250] shock=False  M=2  φ=0.411  loss=1.1883
-[299] SWITCH: M 2 → 0
-[300] shock=False  M=0  φ=0.299  loss=0.7496
-[329] SWITCH: M 0 → 1
---- Shock #2 begins at step 400 ---
-[400] shock=True   M=0  φ=0.581  loss=1.9083
-[419] SWITCH: M 0 → 2
-[425] shock=True   M=2  φ=0.719  loss=1.5779
-[450] shock=True   M=2  φ=0.730  loss=2.4856
-[475] shock=True   M=2  φ=0.640  loss=1.8049
---- End of Shock #2 ---
-[500] shock=False  M=2  φ=0.676  loss=1.7585
----
-🔍 Interpretation
-✔ 1. Unified-LoRA switches correctly under stress
-Enters Multi when φ rises
-Switches to Mirror during both shocks
-Exits Mirror only when E_smooth stabilizes
-✔ 2. Mirror-Lock prevents catastrophic forgetting
-Unlike previous tests (and unlike baseline fixed LoRA), Unified-LoRA:
-Does NOT explode after shocks
-Keeps loss < 2 after both shock exits
-Maintains task performance
-✔ 3. Unified-LoRA recovers smoothly after each shock
-Post-shock recovery:
-Shock #1: 0.897 → 0.749
-Shock #2: 1.577 → 1.758 (stable, no spike)
-This is far better than baseline, which typically jumps to 10+ loss after shocks.
----
-🧠 Why this matters
-This test demonstrates that Unified-LoRA behaves like a true feedback control system:
-It detects instability
-It adjusts its adaptation strategy dynamically
-It protects the base skill during shocks
-It recovers faster and more safely than static LoRA
-This is exactly the kind of robustness needed in:
-Lifelong learning
-Continual fine-tuning
-Noisy or shifting datasets
-Online RLHF loops
----
-🏁 Conclusion
-Unified-LoRA v2, with Mirror-Lock and corrected hysteresis, shows:
-Strong shock robustness
-Low catastrophic forgetting
-Clean mode transitions
-Stable recovery after domain shifts
-These results validate Unified-LoRA as a viable dynamic alternative to traditional LoRA fine-tuning, with potential for real-world deployment.