Spaces:
Sleeping
Sleeping
docs: detailed architecture diagrams with concrete values
Browse files10 visual diagrams (mostly ASCII trees and box layouts) walking through:
1. Three-layer system separation (agent / orchestration / env)
2. Episode timeline (28 steps = 7 days x 4 slots)
3. State tree (hidden profile + visible meters + history with anomalies)
4. What the agent sees (concrete prompt + completion example)
5. Reward stack with worked example: '3 7 5 DEEP_WORK' for seed=42 step 5
- Layer-by-layer reward computation with actual numbers
- Final reward = +1.49 (sum of 4 weighted layers)
6. GRPO update step (8 completions -> advantages -> backprop)
- Shows why mode collapse happens when reward_std=0
7. Dataset structure (replay metadata, NOT supervised labels)
8. Final episode grader (5 components weighted 0.20/0.25/0.15/0.30/0.10)
9. Three eval conditions (discrete-3, in-dist, OOD)
10. Spend & timing per iter
Every diagram uses concrete values from a real example (sampled_42 profile,
specific belief vector, real meter values, real reward calculations).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- docs/architecture.md +497 -0
docs/architecture.md
ADDED
|
@@ -0,0 +1,497 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RhythmEnv β Architecture & Training Flow
|
| 2 |
+
|
| 3 |
+
Visual deep dive into how the env is structured and how the LLM agent
|
| 4 |
+
learns from it. All examples use concrete values (seed=42, sampled
|
| 5 |
+
profile, real numbers from the reward calculation).
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 1. System: three-layer separation
|
| 10 |
+
|
| 11 |
+
```
|
| 12 |
+
ββββββββββββββββββββββββββββββββββββββββββββββ
|
| 13 |
+
β AGENT (Qwen 2.5-3B + LoRA r=16, 4-bit) β
|
| 14 |
+
β β
|
| 15 |
+
β Input: prompt (state + history) β
|
| 16 |
+
β Output: "3 7 5 DEEP_WORK" β
|
| 17 |
+
β β β β β β
|
| 18 |
+
β social|morn|work action β
|
| 19 |
+
β belief|pref|pref β
|
| 20 |
+
βββββββββββββββββββ¬βββββββββββββββββββββββββββ
|
| 21 |
+
β
|
| 22 |
+
β N=8 completions per prompt
|
| 23 |
+
β (sampling temp=1.5)
|
| 24 |
+
βΌ
|
| 25 |
+
βββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 26 |
+
β ORCHESTRATION (TRL GRPOTrainer + Unsloth) β
|
| 27 |
+
β β
|
| 28 |
+
β β’ Picks 1 prompt from dataset (~3000 rows) β
|
| 29 |
+
β β’ Generates 8 completions β
|
| 30 |
+
β β’ Calls 4 reward functions in parallel β
|
| 31 |
+
β β’ Computes group-relative advantages β
|
| 32 |
+
β β’ Backprop on LoRA weights only (~30M) β
|
| 33 |
+
β β’ KL constraint to base Qwen (Ξ²=0.04) β
|
| 34 |
+
βββββββββββββββββββ¬ββββββββββββββββββββββββββββββ
|
| 35 |
+
β
|
| 36 |
+
β env.reset(seed) β step(action)
|
| 37 |
+
β (replay-based reward)
|
| 38 |
+
βΌ
|
| 39 |
+
βββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 40 |
+
β ENVIRONMENT (RhythmEnvironment / FastAPI) β
|
| 41 |
+
β β
|
| 42 |
+
β reset(seed=42) β samples profile β
|
| 43 |
+
β step(action) β updates 5 meters, β
|
| 44 |
+
β returns observation β
|
| 45 |
+
β + per-step reward β
|
| 46 |
+
β β
|
| 47 |
+
β Lives at: huggingface.co/spaces/... β
|
| 48 |
+
βββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
The agent never imports env code. They communicate via the OpenEnv
|
| 52 |
+
HTTP/WebSocket contract: `POST /reset`, `POST /step`, `GET /state`.
|
| 53 |
+
|
| 54 |
+
---
|
| 55 |
+
|
| 56 |
+
## 2. One episode = one week (28 steps)
|
| 57 |
+
|
| 58 |
+
```
|
| 59 |
+
ONE EPISODE = 1 WEEK = 28 STEPS
|
| 60 |
+
|
| 61 |
+
Day 1 (Mon) Day 2 (Tue) ... Day 7 (Sun)
|
| 62 |
+
βββββββ¬ββββββ¬ββββββ¬ββββββ βββββββ¬ββββββ¬ββββββ¬ββββββ βββββββ¬ββββββ¬ββββββ¬ββββββ
|
| 63 |
+
β M β A β E β N β β M β A β E β N β .. β M β A β E β N β
|
| 64 |
+
β 0 β 1 β 2 β 3 β β 4 β 5 β 6 β 7 β β 24 β 25 β 26 β 27 β
|
| 65 |
+
βββββββ΄ββββββ΄ββββββ΄ββββββ βββββββ΄ββββββ΄ββββββ΄ββββββ βββββββ΄ββββββ΄ββββββ΄ββββββ
|
| 66 |
+
β β β
|
| 67 |
+
reset() random event roll (8% per step) done=True
|
| 68 |
+
meters: V=0.7 meditate? deep_work? grader runs
|
| 69 |
+
C=0.7 P=0 S=0.7 sleep? socialize? final_score [0,1]
|
| 70 |
+
Cn=0.5
|
| 71 |
+
|
| 72 |
+
At each step the agent picks 1 of 10 actions:
|
| 73 |
+
|
| 74 |
+
ββββββββββββββββββ¬ββββββββββββββββββ¬ββββββββββββββββ¬βββββββββββββββββ
|
| 75 |
+
β PRODUCTIVITY β RECOVERY β SOCIAL β LEISURE β
|
| 76 |
+
ββββββββββββββββββΌββββββββββββββββββΌββββββββββββββββΌβββββββββββββββββ€
|
| 77 |
+
β DEEP_WORK β SLEEP β FAMILY_TIME β ME_TIME β
|
| 78 |
+
β ADMIN_WORK β EXERCISE β SOCIALIZE β BINGE_WATCH β
|
| 79 |
+
β LEARN β MEDITATE β β β
|
| 80 |
+
ββββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββββββββ΄βββββββββββββββββ
|
| 81 |
+
|
| 82 |
+
Time-of-day multipliers (apply to action effects):
|
| 83 |
+
Morning (M): cognition gains Γ 1.2 vitality drain Γ 0.8
|
| 84 |
+
Afternoon (A): cognition gains Γ 1.0 vitality drain Γ 1.0
|
| 85 |
+
Evening (E): cognition gains Γ 0.8 vitality drain Γ 1.1
|
| 86 |
+
Night (N): cognition gains Γ 0.6 vitality drain Γ 1.3
|
| 87 |
+
(sleep BYPASSES these multipliers)
|
| 88 |
+
|
| 89 |
+
Critical thresholds:
|
| 90 |
+
Any meter < 0.1 at end of step β -0.30 reward penalty (one per meter)
|
| 91 |
+
Connection decays passively every step (-0.01 to -0.02 per profile)
|
| 92 |
+
```
|
| 93 |
+
|
| 94 |
+
---
|
| 95 |
+
|
| 96 |
+
## 3. State tree: what the env tracks (and what's hidden)
|
| 97 |
+
|
| 98 |
+
```
|
| 99 |
+
RhythmEnvironment instance state
|
| 100 |
+
β
|
| 101 |
+
βββ _profile βββββ HIDDEN from agent ββββββββββ
|
| 102 |
+
β β β
|
| 103 |
+
β βββ name: "sampled_42" β
|
| 104 |
+
β βββ reward_weights: β
|
| 105 |
+
β β {V: 0.05, C: 0.05, P: 0.30, β
|
| 106 |
+
β β S: 0.20, Cn: 0.40} β
|
| 107 |
+
β β β Used INTERNALLY
|
| 108 |
+
β βββ social_vitality_multiplier: 1.8 β to compute reward
|
| 109 |
+
β βββ morning_cognition_bonus: 1.5 β and modify action
|
| 110 |
+
β βββ evening_night_cognition_bonus: None β effects.
|
| 111 |
+
β βββ morning_penalty: None β Belief reward
|
| 112 |
+
β βββ binge_shame: False β rewards INFERRING
|
| 113 |
+
β βββ progress_serenity_bonus: 0.04 β this profile.
|
| 114 |
+
β βββ idle_serenity_decay: 0.02 β
|
| 115 |
+
β βββ vitality_decay_rate: 0.01 β
|
| 116 |
+
β βββ stress_tolerance: 0.22 β
|
| 117 |
+
β βββ event_impact_multiplier: 0.7 β
|
| 118 |
+
β βββ connection_decay_rate: 0.012 β
|
| 119 |
+
β βββ solo_serenity_bonus: 0.05 β
|
| 120 |
+
β βββ social_connection_multiplier: 1.4 β
|
| 121 |
+
β βββ social_serenity_bonus: 0.03 β
|
| 122 |
+
β βββ work_vitality_recovery: 0.03 β
|
| 123 |
+
β βββ (14 hidden parameters total) β
|
| 124 |
+
β βββββββββββββββββββββββββββββββββββββββββββ
|
| 125 |
+
β
|
| 126 |
+
β Reduces to a 3-D BELIEF VECTOR (the inference target):
|
| 127 |
+
β profile_to_belief_vector(profile) β [0.30, 0.70, 0.50]
|
| 128 |
+
β β β β
|
| 129 |
+
β β β work_pref
|
| 130 |
+
β β morning_pref
|
| 131 |
+
β social_pref
|
| 132 |
+
β
|
| 133 |
+
βββ meters βββββββ visible to agent βββββββββββββββββββββββββββββ
|
| 134 |
+
β βββ _vitality: 0.62 (range 0β1) β
|
| 135 |
+
β βββ _cognition: 0.51 β
|
| 136 |
+
β βββ _progress: 0.24 β Agent
|
| 137 |
+
β βββ _serenity: 0.71 β observes
|
| 138 |
+
β βββ _connection: 0.38 β in prompt
|
| 139 |
+
β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 140 |
+
β
|
| 141 |
+
βββ _timestep: 5
|
| 142 |
+
β
|
| 143 |
+
βββ _step_history: list[StepRecord] βββ visible to agent βββββββ
|
| 144 |
+
β βββ (last 7 steps) β
|
| 145 |
+
β β step 0: deep_work β +0.42 β
|
| 146 |
+
β β deltas: V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00 β
|
| 147 |
+
β β anomalies: V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00 ββ profile fingerprint
|
| 148 |
+
β β step 1: sleep β +0.18 β
|
| 149 |
+
β β deltas: V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00 β
|
| 150 |
+
β β anomalies: V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00 β
|
| 151 |
+
β β ... β
|
| 152 |
+
β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 153 |
+
β
|
| 154 |
+
βββ _step_rewards: [+0.42, +0.18, +0.31, +0.55, ...]
|
| 155 |
+
βββ used by grader to compute adaptation_score (late-half - early-half)
|
| 156 |
+
```
|
| 157 |
+
|
| 158 |
+
---
|
| 159 |
+
|
| 160 |
+
## 4. What the agent sees (concrete prompt)
|
| 161 |
+
|
| 162 |
+
```
|
| 163 |
+
ββββββββββββββββββββββββββββββββββββββ SYSTEM PROMPT ββββββββββββββββββββββββββββββββββ
|
| 164 |
+
β You are a life-management agent helping a person whose preferences are HIDDEN. β
|
| 165 |
+
β Each step, output ONE LINE in this exact format: β
|
| 166 |
+
β S M W ACTION_NAME β
|
| 167 |
+
β S = social pref (0=hates, 9=loves), M = morning, W = work β
|
| 168 |
+
β ACTION_NAME β {DEEP_WORK, ADMIN_WORK, LEARN, SLEEP, EXERCISE, MEDITATE, β
|
| 169 |
+
β FAMILY_TIME, SOCIALIZE, ME_TIME, BINGE_WATCH} β
|
| 170 |
+
β Example: 3 8 7 DEEP_WORK β
|
| 171 |
+
β Tactics: probe early, exploit late; don't repeat actions; ... β
|
| 172 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 173 |
+
βββββββββββββββββββββββββββββββββββββ USER PROMPT βββββββββββββββββββββββββββββββββββββ
|
| 174 |
+
β Step: 5/28 (Tuesday Afternoon) β
|
| 175 |
+
β Remaining steps: 22 β
|
| 176 |
+
β β
|
| 177 |
+
β Meters: β
|
| 178 |
+
β Vitality: 0.62 β
|
| 179 |
+
β Cognition: 0.51 β
|
| 180 |
+
β Progress: 0.24 β
|
| 181 |
+
β Serenity: 0.71 β
|
| 182 |
+
β Connection: 0.38 β
|
| 183 |
+
β β
|
| 184 |
+
β Recent history (anom = how this person deviated from neutral baseline): β
|
| 185 |
+
β step 0: deep_work -> reward +0.42 (V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00) β
|
| 186 |
+
β [anom V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00] β
|
| 187 |
+
β step 1: sleep -> reward +0.18 (V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00) β
|
| 188 |
+
β [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] β
|
| 189 |
+
β step 2: socialize -> reward -0.05 (V-0.11 C-0.03 P+0.00 S+0.04 Cn+0.17) β
|
| 190 |
+
β [anom V-0.05 C+0.00 P+0.00 S+0.00 Cn+0.05] β strong profile signal β
|
| 191 |
+
β step 3: meditate -> reward +0.30 (V+0.03 C+0.08 P+0.00 S+0.20 Cn+0.00) β
|
| 192 |
+
β [anom V+0.00 C+0.00 P+0.00 S+0.05 Cn+0.00] β solo bonus visible β
|
| 193 |
+
β step 4: deep_work -> reward +0.18 (V-0.06 C-0.06 P+0.18 S+0.04 Cn+0.00) β
|
| 194 |
+
β [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] β
|
| 195 |
+
β β
|
| 196 |
+
β Output your belief, then your action (format: S M W ACTION_NAME): β
|
| 197 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 198 |
+
|
| 199 |
+
LLM completion (1 of 8 sampled at temp=1.5):
|
| 200 |
+
ββββββββββββββββββββββββ
|
| 201 |
+
β "3 7 5 DEEP_WORK" β
|
| 202 |
+
ββββββββββββ¬ββββββββββββ
|
| 203 |
+
βΌ
|
| 204 |
+
Parsed: belief = [0.33, 0.78, 0.56]
|
| 205 |
+
action = DEEP_WORK
|
| 206 |
+
```
|
| 207 |
+
|
| 208 |
+
---
|
| 209 |
+
|
| 210 |
+
## 5. The reward stack (4 layers, with concrete values for the example above)
|
| 211 |
+
|
| 212 |
+
```
|
| 213 |
+
βββββββββββββββββββββββββββββββββββββββ
|
| 214 |
+
β Completion: "3 7 5 DEEP_WORK" β
|
| 215 |
+
β for prompt above (seed=42, step 5) β
|
| 216 |
+
βββββββββββββββββββ¬ββββββββββββββββββββ
|
| 217 |
+
β
|
| 218 |
+
ββββββββββββββββββββββββββββββββΌββββββββββββββββββββββββββββββοΏ½οΏ½οΏ½
|
| 219 |
+
β β β
|
| 220 |
+
βΌ βΌ βΌ
|
| 221 |
+
ββββββββββββββββ ββββββββββββββββββββ ββββββββββββββββββββββ
|
| 222 |
+
β format_valid β β action_legal β β env_reward β
|
| 223 |
+
ββββββββββββββββ€ ββββββββββββββββββββ€ ββββββββββββββββββββββ€
|
| 224 |
+
β parses? β β action β 10 ? β β env.replay(seed=42,β
|
| 225 |
+
β has belief? β β β β step5, action) β
|
| 226 |
+
β β β DEEP_WORK β β β β
|
| 227 |
+
β β +1.0 β β reward 0.0 β β deltas: β
|
| 228 |
+
β Γ wt 0.05 β β Γ wt 0.05 β β V-0.10 C-0.12 β
|
| 229 |
+
β β β β β P+0.18 S-0.05 β
|
| 230 |
+
β β +0.05 β β β 0.00 β β Cn+0.00 β
|
| 231 |
+
ββββββββββββββββ ββββββββββββββββββββ β β
|
| 232 |
+
β profile_reward = β
|
| 233 |
+
β sum(deltas Γ β
|
| 234 |
+
β prof_weights) β
|
| 235 |
+
β Γ 15 β
|
| 236 |
+
β = (-0.10Γ0.05 β
|
| 237 |
+
β + -0.12Γ0.05 β
|
| 238 |
+
β + 0.18Γ0.30 β
|
| 239 |
+
β + -0.05Γ0.20 β
|
| 240 |
+
β + 0Γ0.40) Γ 15 β
|
| 241 |
+
β = +0.65 β
|
| 242 |
+
β β
|
| 243 |
+
β + grader_bias: β
|
| 244 |
+
β 0.5Γ0.18+0.4Γ0 β
|
| 245 |
+
β = +0.09 β
|
| 246 |
+
β + new-action: +0.07β
|
| 247 |
+
β + b-act coupling: β
|
| 248 |
+
β work=0.56 (mid), β
|
| 249 |
+
β no bonus = 0 β
|
| 250 |
+
β - cycle pen: 0 β
|
| 251 |
+
β - rep pen: 0 β
|
| 252 |
+
β β
|
| 253 |
+
β env_reward = +0.81 β
|
| 254 |
+
β Γ wt 1.5 = +1.21 β
|
| 255 |
+
ββββββββββββββββββββββ
|
| 256 |
+
β
|
| 257 |
+
β
|
| 258 |
+
βΌ
|
| 259 |
+
βββββββββββββββββββββββββββββββ
|
| 260 |
+
β belief_accuracy β
|
| 261 |
+
βββββββββββββββββββββββββββββββ€
|
| 262 |
+
β true belief (sampled_42) β
|
| 263 |
+
β = [0.30, 0.70, 0.50] β
|
| 264 |
+
β agent belief β
|
| 265 |
+
β = [0.33, 0.78, 0.56] β
|
| 266 |
+
β β
|
| 267 |
+
β MAE = (0.03 + 0.08 + 0.06) β
|
| 268 |
+
β / 3 = 0.057 β
|
| 269 |
+
β similarity = 0.943 β
|
| 270 |
+
β β
|
| 271 |
+
β baseline (constant 0.5): β
|
| 272 |
+
β MAE = (0.20+0.20+0.00)/3 β
|
| 273 |
+
β = 0.133 β
|
| 274 |
+
β similarity = 0.867 β
|
| 275 |
+
β β
|
| 276 |
+
β reward = 0.943 - 0.867 β
|
| 277 |
+
β = +0.076 β
|
| 278 |
+
β Γ wt 3.0 = +0.23 β
|
| 279 |
+
βββββββββββββββββββββββββββββββ
|
| 280 |
+
|
| 281 |
+
Ξ£ TOTAL REWARD (this completion)
|
| 282 |
+
= 0.05 + 0.00 + 1.21 + 0.23
|
| 283 |
+
= +1.49
|
| 284 |
+
```
|
| 285 |
+
|
| 286 |
+
---
|
| 287 |
+
|
| 288 |
+
## 6. GRPO update step (8 completions β 1 gradient update)
|
| 289 |
+
|
| 290 |
+
```
|
| 291 |
+
ONE TRAINING STEP (one row from the dataset β one gradient update)
|
| 292 |
+
|
| 293 |
+
ββ pick prompt βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 294 |
+
β prompt#1247 from dataset: β
|
| 295 |
+
β state: step 5, seed=42, history=[deep_work, sleep, ...] β
|
| 296 |
+
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 297 |
+
β
|
| 298 |
+
βΌ generate 8 completions @ temp=1.5
|
| 299 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 300 |
+
β c1: "3 7 5 DEEP_WORK" β reward +1.49 β
|
| 301 |
+
β c2: "5 5 5 SLEEP" β reward -0.21 (constant belief: -0.03 β
|
| 302 |
+
β + sleep replay: -0.18) β
|
| 303 |
+
β c3: "3 6 4 ADMIN_WORK" β reward +1.10 β
|
| 304 |
+
β c4: "4 7 6 LEARN" β reward +1.32 β
|
| 305 |
+
β c5: "2 8 3 MEDITATE" β reward +0.45 (rep pen: -0.10 since β
|
| 306 |
+
β meditate at step 3 too) β
|
| 307 |
+
β c6: "3 7 7 DEEP_WORK" β reward +1.55 (belief slightly better) β
|
| 308 |
+
β c7: "5 5 5 EXERCISE" β reward -0.15 β
|
| 309 |
+
β c8: "4 6 5 FAMILY_TIME" β reward +0.92 β
|
| 310 |
+
β β
|
| 311 |
+
β group_mean = +0.81 β
|
| 312 |
+
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 313 |
+
β
|
| 314 |
+
βΌ advantages = reward - group_mean
|
| 315 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 316 |
+
β ADVANTAGE (this is what GRPO actually backprops on) β
|
| 317 |
+
β c1: +0.68 β strongly preferred β
|
| 318 |
+
β c2: -1.02 β strongly discouraged (constant belief) β
|
| 319 |
+
β c3: +0.29 β
|
| 320 |
+
β c4: +0.51 β
|
| 321 |
+
β c5: -0.36 β
|
| 322 |
+
β c6: +0.74 β most preferred β
|
| 323 |
+
β c7: -0.96 β
|
| 324 |
+
β c8: +0.11 β
|
| 325 |
+
β β
|
| 326 |
+
β KEY INSIGHT: only the SPREAD matters. If all 8 had the β
|
| 327 |
+
β same reward, advantages would all be 0 β no gradient. β
|
| 328 |
+
β This is why iter 1 mode-collapsed: format_valid +1.0 for β
|
| 329 |
+
β every completion meant zero variance from that layer. β
|
| 330 |
+
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 331 |
+
β
|
| 332 |
+
βΌ policy loss = -E[ adv Γ log_prob(completion) ]
|
| 333 |
+
β + Ξ² Γ KL(policy || base_qwen)
|
| 334 |
+
β
|
| 335 |
+
βΌ backprop (only LoRA weights, ~30M params)
|
| 336 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 337 |
+
β Model nudge: β
|
| 338 |
+
β push: "3 7 5 DEEP_WORK"-like outputs UP β
|
| 339 |
+
β push: "3 7 7 DEEP_WORK"-like outputs UP β
|
| 340 |
+
β pull: "5 5 5 *"-like outputs DOWN β
|
| 341 |
+
β β
|
| 342 |
+
β KL constraint: prevents the policy from diverging too β
|
| 343 |
+
β far from base Qwen (avoids gibberish drift). β
|
| 344 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 345 |
+
|
| 346 |
+
Repeat 800-2000 times. Each step ~3-5 sec on A100.
|
| 347 |
+
```
|
| 348 |
+
|
| 349 |
+
---
|
| 350 |
+
|
| 351 |
+
## 7. The dataset is just starting positions (not labels)
|
| 352 |
+
|
| 353 |
+
```
|
| 354 |
+
DATASET (~3000 rows, generated ONCE before training)
|
| 355 |
+
|
| 356 |
+
For 200-300 episodes:
|
| 357 |
+
env.reset(seed=N)
|
| 358 |
+
for step in range(28):
|
| 359 |
+
record {
|
| 360 |
+
prompt: [system, user_for_this_state],
|
| 361 |
+
seed: N, ββββ replay metadata
|
| 362 |
+
step_index: <current step>,
|
| 363 |
+
action_history: <actions taken so far>,
|
| 364 |
+
profile_mode: "continuous",
|
| 365 |
+
}
|
| 366 |
+
env.step(rollout_policy(obs)) ββ rollout=heuristic OR random
|
| 367 |
+
(only matters for STATE diversity,
|
| 368 |
+
the agent's training generations
|
| 369 |
+
REPLACE these actions)
|
| 370 |
+
|
| 371 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 372 |
+
β A row from the dataset: β
|
| 373 |
+
β { β
|
| 374 |
+
β prompt: [...full state at step 5 of episode seed=42...] β
|
| 375 |
+
β seed: 42, β
|
| 376 |
+
β step_index: 5, β
|
| 377 |
+
β action_history: ["deep_work", "sleep", "socialize", β
|
| 378 |
+
β "meditate", "deep_work"], β
|
| 379 |
+
β profile_mode: "continuous" β
|
| 380 |
+
β } β
|
| 381 |
+
β β
|
| 382 |
+
β NOTE: NO "correct action" or "label" anywhere. β
|
| 383 |
+
β The reward function reconstructs the env from this metadata β
|
| 384 |
+
β and scores whatever action the LLM picks. β
|
| 385 |
+
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 386 |
+
|
| 387 |
+
This is fundamentally different from supervised learning:
|
| 388 |
+
- Supervised: (input, target_output) β model learns to mimic target
|
| 389 |
+
- GRPO: (prompt, replay_metadata) β model learns to maximize reward
|
| 390 |
+
```
|
| 391 |
+
|
| 392 |
+
---
|
| 393 |
+
|
| 394 |
+
## 8. Final episode grader (only fires at step 28)
|
| 395 |
+
|
| 396 |
+
```
|
| 397 |
+
_grade_episode() β runs at done=True
|
| 398 |
+
|
| 399 |
+
ββββββββββββββββ
|
| 400 |
+
β final_score β
|
| 401 |
+
β β [0, 1] β
|
| 402 |
+
ββββββββ¬ββββββββ
|
| 403 |
+
ββββββββββββββββββββ¬ββββββββββΌβββββββββββββββ¬βββββββββββββββ
|
| 404 |
+
β β β β β
|
| 405 |
+
βΌ βΌ βΌ βΌ βΌ
|
| 406 |
+
ββββββββββββββ ββββββββββββββ ββββββββββββ ββββββββββββββ ββββββββββββββ
|
| 407 |
+
β crash_free β β progress β β connectionβ β adaptation β β efficiency β
|
| 408 |
+
β Γ 0.20 β β Γ 0.25 β β Γ 0.15 β β Γ 0.30 β β Γ 0.10 β
|
| 409 |
+
ββββββββββββββ€ ββββββββββββββ€ ββββββββββββ€ ββββββββββββββ€ ββββββββββββββ€
|
| 410 |
+
β 1 - crashesβ β final P β β final Cn β β late-half β β avg_reward β
|
| 411 |
+
β /total_ck β β value β β value β β mean rewardβ β normalized β
|
| 412 |
+
β β β β β β β - early β β to [0,1] β
|
| 413 |
+
β e.g. 0.95 β β e.g. 0.42 β β e.g. 0.51β β e.g. +0.18 β β e.g. 0.55 β
|
| 414 |
+
β Γ0.20=0.19 β β Γ0.25=0.105β β Γ0.15=0.08β β Γ0.30=0.054β β Γ0.10=0.055β
|
| 415 |
+
ββββββββββββββ ββββββββββββββ βββββββββββοΏ½οΏ½ ββββββββββββββ ββββββββββββββ
|
| 416 |
+
|
| 417 |
+
Ξ£ = 0.19 + 0.105 + 0.08 + 0.054 + 0.055
|
| 418 |
+
= 0.484 β final_score
|
| 419 |
+
|
| 420 |
+
Plus iter 4 sparse terminal reward (added to step 27's per-step reward):
|
| 421 |
+
terminal_bonus = (0.484 - 0.5) Γ 5 = -0.08
|
| 422 |
+
|
| 423 |
+
This means: at step 27, agent gets last per-step reward + bonus from grader.
|
| 424 |
+
This is the only direct gradient signal pointing at the actual episode quality.
|
| 425 |
+
```
|
| 426 |
+
|
| 427 |
+
---
|
| 428 |
+
|
| 429 |
+
## 9. Three eval conditions (post-training)
|
| 430 |
+
|
| 431 |
+
```
|
| 432 |
+
inference_eval.py runs ALL THREE
|
| 433 |
+
|
| 434 |
+
ββββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββ
|
| 435 |
+
β discrete-3-profiles β continuous-in-distributionβ continuous-OOD β
|
| 436 |
+
β (legacy comparison) β (was the agent able to β (does meta-policy β
|
| 437 |
+
β β learn the meta-policy?) β generalize?) β
|
| 438 |
+
ββββββββββββββββββββββββββββββΌβββββββββββββββββββββββββββββΌβββββββββββββββββββββββββββββ€
|
| 439 |
+
β env.reset(seed=N, β env.reset(seed=N) β env.reset(seed=10000+N) β
|
| 440 |
+
β profile="introvert_ β β samples from training β β samples from a region β
|
| 441 |
+
β morning") β distribution β never seen in trainingβ
|
| 442 |
+
β β β β
|
| 443 |
+
β β 3 hardcoded profiles β β ~10 sampled profiles β β ~10 sampled profiles β
|
| 444 |
+
β (from PROFILES list) β from seeds 100..110 β from seeds 10000..10010 β
|
| 445 |
+
β β β β
|
| 446 |
+
β Each strategy plays each β Each strategy plays each β Each strategy plays each β
|
| 447 |
+
β profile 5x = 15 episodes β seed 1x = 10 episodes β seed 1x = 10 episodes β
|
| 448 |
+
β β β β
|
| 449 |
+
β Strategies tested: β Strategies tested: β Strategies tested: β
|
| 450 |
+
β β’ random β β’ random β β’ random β
|
| 451 |
+
β β’ heuristic β β’ heuristic β β’ heuristic β
|
| 452 |
+
β β’ model (trained Qwen) β β’ model (trained Qwen) β β’ model (trained Qwen) β
|
| 453 |
+
ββββββββββββββββββββββββββββββ΄βββββββββββββββββββββββββββββ΄βββββββββββββββββββββββββββββ
|
| 454 |
+
|
| 455 |
+
THE KEY METRIC: trained model's score on continuous-OOD vs heuristic baseline.
|
| 456 |
+
|
| 457 |
+
Heuristic baseline (profile-blind hand rules): score 0.580 on OOD.
|
| 458 |
+
Trained meta-RL agent's target: > 0.580 on OOD.
|
| 459 |
+
|
| 460 |
+
If the trained agent beats the heuristic on OOD (profiles never seen in
|
| 461 |
+
training), that's direct proof it learned the SKILL of profile inference,
|
| 462 |
+
not just memorized training profiles.
|
| 463 |
+
```
|
| 464 |
+
|
| 465 |
+
---
|
| 466 |
+
|
| 467 |
+
## 10. Spend & timing (concrete)
|
| 468 |
+
|
| 469 |
+
```
|
| 470 |
+
HARDWARE: A100 large (80GB) on HF Jobs at $2.50/hr ($0.0417/min)
|
| 471 |
+
|
| 472 |
+
FAST_MODE (200-800 steps):
|
| 473 |
+
dataset gen: ~2 min
|
| 474 |
+
model load: ~3 min
|
| 475 |
+
training: ~10-25 min (depends on steps)
|
| 476 |
+
eval: ~3 min
|
| 477 |
+
plot + upload: ~2 min
|
| 478 |
+
ββββββββββββββββββββββββββββ
|
| 479 |
+
total: 20-35 min ($0.80-1.50 per iter)
|
| 480 |
+
|
| 481 |
+
FULL RUN (2000 steps):
|
| 482 |
+
dataset gen: ~3 min
|
| 483 |
+
model load: ~3 min
|
| 484 |
+
training: ~60-90 min
|
| 485 |
+
eval: ~3 min
|
| 486 |
+
plot + upload: ~2 min
|
| 487 |
+
ββββββββββββββββββββββββββββ
|
| 488 |
+
total: 70-100 min ($3-4)
|
| 489 |
+
|
| 490 |
+
Iter 1 (200 steps): $0.50 β mode collapse (single action)
|
| 491 |
+
Iter 2 (400 steps): $1.50 β mode collapse (2-cycle)
|
| 492 |
+
Iter 3 (800 steps): $5 β³ in flight (control)
|
| 493 |
+
Iter 4 (800 steps): $5 β³ in flight (with full fixes)
|
| 494 |
+
Final (2000 steps): $4 β³ pending iter 3+4 results
|
| 495 |
+
ββββββ
|
| 496 |
+
~$16 of $30 budget
|
| 497 |
+
```
|