Spaces:
Sleeping
Sleeping
RhythmEnv β Architecture & Training Flow
Visual deep dive into how the env is structured and how the LLM agent learns from it. All examples use concrete values (seed=42, sampled profile, real numbers from the reward calculation).
1. System: three-layer separation
ββββββββββββββββββββββββββββββββββββββββββββββ
β AGENT (Qwen 2.5-3B + LoRA r=8, 4-bit) β
β β
β Input: prompt (state + history) β
β Output: "3 7 5 DEEP_WORK" β
β β β β β β
β social|morn|work action β
β belief|pref|pref β
βββββββββββββββββββ¬βββββββββββββββββββββββββββ
β
β N=8 completions per prompt
β (sampling temp=1.5)
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββ
β ORCHESTRATION (TRL GRPOTrainer + Unsloth) β
β β
β β’ Picks 1 prompt from dataset (~3000 rows) β
β β’ Generates 8 completions β
β β’ Calls 4 reward functions in parallel β
β β’ Computes group-relative advantages β
β β’ Backprop on LoRA weights only (~30M) β
β β’ KL constraint to base Qwen (Ξ²=0.04) β
βββββββββββββββββββ¬ββββββββββββββββββββββββββββββ
β
β env.reset(seed) β step(action)
β (replay-based reward)
βΌ
βββββββββββββββββββββββββββββββββββββββββββββββββ
β ENVIRONMENT (RhythmEnvironment / FastAPI) β
β β
β reset(seed=42) β samples profile β
β step(action) β updates 5 meters, β
β returns observation β
β + per-step reward β
β β
β Lives at: huggingface.co/spaces/... β
βββββββββββββββββββββββββββββββββββββββββββββββββ
The agent never imports env code. They communicate via the OpenEnv
HTTP/WebSocket contract: POST /reset, POST /step, GET /state.
2. One episode = one week (28 steps)
ONE EPISODE = 1 WEEK = 28 STEPS
Day 1 (Mon) Day 2 (Tue) ... Day 7 (Sun)
βββββββ¬ββββββ¬ββββββ¬ββββββ βββββββ¬ββββββ¬ββββββ¬ββββββ βββββββ¬ββββββ¬ββββββ¬ββββββ
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
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
βββββββ΄ββββββ΄ββββββ΄ββββββ βββββββ΄ββββββ΄ββββββ΄ββββββ βββββββ΄ββββββ΄ββββββ΄ββββββ
β β β
reset() random event roll (8% per step) done=True
meters: V=0.7 meditate? deep_work? grader runs
C=0.7 P=0 S=0.7 sleep? socialize? final_score [0,1]
Cn=0.5
At each step the agent picks 1 of 10 actions:
ββββββββββββββββββ¬ββββββββββββββββββ¬ββββββββββββββββ¬βββββββββββββββββ
β PRODUCTIVITY β RECOVERY β SOCIAL β LEISURE β
ββββββββββββββββββΌββββββββββββββββββΌββββββββββββββββΌβββββββββββββββββ€
β DEEP_WORK β SLEEP β FAMILY_TIME β ME_TIME β
β ADMIN_WORK β EXERCISE β SOCIALIZE β BINGE_WATCH β
β LEARN β MEDITATE β β β
ββββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββββββββ΄βββββββββββββββββ
Time-of-day multipliers (apply to action effects):
Morning (M): cognition gains Γ 1.2 vitality drain Γ 0.8
Afternoon (A): cognition gains Γ 1.0 vitality drain Γ 1.0
Evening (E): cognition gains Γ 0.8 vitality drain Γ 1.1
Night (N): cognition gains Γ 0.6 vitality drain Γ 1.3
(sleep BYPASSES these multipliers)
Critical thresholds:
Any meter < 0.1 at end of step β -0.30 reward penalty (one per meter)
Connection decays passively every step (-0.01 to -0.02 per profile)
3. State tree: what the env tracks (and what's hidden)
RhythmEnvironment instance state
β
βββ _profile βββββ HIDDEN from agent ββββββββββ
β β β
β βββ name: "sampled_42" β
β βββ reward_weights: β
β β {V: 0.05, C: 0.05, P: 0.30, β
β β S: 0.20, Cn: 0.40} β
β β β Used INTERNALLY
β βββ social_vitality_multiplier: 1.8 β to compute reward
β βββ morning_cognition_bonus: 1.5 β and modify action
β βββ evening_night_cognition_bonus: None β effects.
β βββ morning_penalty: None β Belief reward
β βββ binge_shame: False β rewards INFERRING
β βββ progress_serenity_bonus: 0.04 β this profile.
β βββ idle_serenity_decay: 0.02 β
β βββ vitality_decay_rate: 0.01 β
β βββ stress_tolerance: 0.22 β
β βββ event_impact_multiplier: 0.7 β
β βββ connection_decay_rate: 0.012 β
β βββ solo_serenity_bonus: 0.05 β
β βββ social_connection_multiplier: 1.4 β
β βββ social_serenity_bonus: 0.03 β
β βββ work_vitality_recovery: 0.03 β
β βββ (14 hidden parameters total) β
β βββββββββββββββββββββββββββββββββββββββββββ
β
β Reduces to a 3-D BELIEF VECTOR (the inference target):
β profile_to_belief_vector(profile) β [0.30, 0.70, 0.50]
β β β β
β β β work_pref
β β morning_pref
β social_pref
β
βββ meters βββββββ visible to agent βββββββββββββββββββββββββββββ
β βββ _vitality: 0.62 (range 0β1) β
β βββ _cognition: 0.51 β
β βββ _progress: 0.24 β Agent
β βββ _serenity: 0.71 β observes
β βββ _connection: 0.38 β in prompt
β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
βββ _timestep: 5
β
βββ _step_history: list[StepRecord] βββ visible to agent βββββββ
β βββ (last 7 steps) β
β β step 0: deep_work β +0.42 β
β β deltas: V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00 β
β β anomalies: V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00 ββ profile fingerprint
β β step 1: sleep β +0.18 β
β β deltas: V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00 β
β β anomalies: V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00 β
β β ... β
β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
βββ _step_rewards: [+0.42, +0.18, +0.31, +0.55, ...]
βββ used by grader to compute adaptation_score (late-half - early-half)
4. What the agent sees (concrete prompt)
ββββββββββββββββββββββββββββββββββββββ SYSTEM PROMPT ββββββββββββββββββββββββββββββββββ
β You are a life-management agent helping a person whose preferences are HIDDEN. β
β Each step, output ONE LINE in this exact format: β
β S M W ACTION_NAME β
β S = social pref (0=hates, 9=loves), M = morning, W = work β
β ACTION_NAME β {DEEP_WORK, ADMIN_WORK, LEARN, SLEEP, EXERCISE, MEDITATE, β
β FAMILY_TIME, SOCIALIZE, ME_TIME, BINGE_WATCH} β
β Example: 3 8 7 DEEP_WORK β
β Tactics: probe early, exploit late; don't repeat actions; ... β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
βββββββββββββββββββββββββββββββββββββ USER PROMPT βββββββββββββββββββββββββββββββββββββ
β Step: 5/28 (Tuesday Afternoon) β
β Remaining steps: 22 β
β β
β Meters: β
β Vitality: 0.62 β
β Cognition: 0.51 β
β Progress: 0.24 β
β Serenity: 0.71 β
β Connection: 0.38 β
β β
β Recent history (anom = how this person deviated from neutral baseline): β
β step 0: deep_work -> reward +0.42 (V-0.10 C-0.12 P+0.18 S-0.05 Cn+0.00) β
β [anom V+0.00 C+0.00 P+0.06 S+0.00 Cn+0.00] β
β step 1: sleep -> reward +0.18 (V+0.20 C+0.10 P+0.00 S+0.05 Cn+0.00) β
β [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] β
β step 2: socialize -> reward -0.05 (V-0.11 C-0.03 P+0.00 S+0.04 Cn+0.17) β
β [anom V-0.05 C+0.00 P+0.00 S+0.00 Cn+0.05] β strong profile signal β
β step 3: meditate -> reward +0.30 (V+0.03 C+0.08 P+0.00 S+0.20 Cn+0.00) β
β [anom V+0.00 C+0.00 P+0.00 S+0.05 Cn+0.00] β solo bonus visible β
β step 4: deep_work -> reward +0.18 (V-0.06 C-0.06 P+0.18 S+0.04 Cn+0.00) β
β [anom V+0.00 C+0.00 P+0.00 S+0.00 Cn+0.00] β
β β
β Output your belief, then your action (format: S M W ACTION_NAME): β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
LLM completion (1 of 8 sampled at temp=1.5):
ββββββββββββββββββββββββ
β "3 7 5 DEEP_WORK" β
ββββββββββββ¬ββββββββββββ
βΌ
Parsed: belief = [0.33, 0.78, 0.56]
action = DEEP_WORK
5. The reward stack (4 layers, with concrete values for the example above)
βββββββββββββββββββββββββββββββββββββββ
β Completion: "3 7 5 DEEP_WORK" β
β for prompt above (seed=42, step 5) β
βββββββββββββββββββ¬ββββββββββββββββββββ
β
ββββββββββββββββββββββββββββββββΌβββββββββββββββββββββββββββββββ
β β β
βΌ βΌ βΌ
ββββββββββββββββ ββββββββββββββββββββ ββββββββββββββββββββββ
β format_valid β β action_legal β β env_reward β
ββββββββββββββββ€ ββββββββββββββββββββ€ ββββββββββββββββββββββ€
β parses? β β action β 10 ? β β env.replay(seed=42,β
β has belief? β β β β step5, action) β
β β β DEEP_WORK β β β β
β β +1.0 β β reward 0.0 β β deltas: β
β Γ wt 0.05 β β Γ wt 0.05 β β V-0.10 C-0.12 β
β β β β β P+0.18 S-0.05 β
β β +0.05 β β β 0.00 β β Cn+0.00 β
ββββββββββββββββ ββββββββββββββββββββ β β
β profile_reward = β
β sum(deltas Γ β
β prof_weights) β
β Γ 15 β
β = (-0.10Γ0.05 β
β + -0.12Γ0.05 β
β + 0.18Γ0.30 β
β + -0.05Γ0.20 β
β + 0Γ0.40) Γ 15 β
β = +0.65 β
β β
β + grader_bias: β
β 0.5Γ0.18+0.4Γ0 β
β = +0.09 β
β + new-action: +0.07β
β + b-act coupling: β
β work=0.56 (mid), β
β no bonus = 0 β
β - cycle pen: 0 β
β - rep pen: 0 β
β β
β env_reward = +0.81 β
β Γ wt 1.5 = +1.21 β
ββββββββββββββββββββββ
β
β
βΌ
βββββββββββββββββββββββββββββββ
β belief_accuracy β
βββββββββββββββββββββββββββββββ€
β true belief (sampled_42) β
β = [0.30, 0.70, 0.50] β
β agent belief β
β = [0.33, 0.78, 0.56] β
β β
β MAE = (0.03 + 0.08 + 0.06) β
β / 3 = 0.057 β
β similarity = 0.943 β
β β
β baseline (constant 0.5): β
β MAE = (0.20+0.20+0.00)/3 β
β = 0.133 β
β similarity = 0.867 β
β β
β reward = 0.943 - 0.867 β
β = +0.076 β
β Γ wt 3.0 = +0.23 β
βββββββββββββββββββββββββββββββ
Ξ£ TOTAL REWARD (this completion)
= 0.05 + 0.00 + 1.21 + 0.23
= +1.49
6. GRPO update step (8 completions β 1 gradient update)
ONE TRAINING STEP (one row from the dataset β one gradient update)
ββ pick prompt βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β prompt#1247 from dataset: β
β state: step 5, seed=42, history=[deep_work, sleep, ...] β
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
βΌ generate 8 completions @ temp=1.5
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β c1: "3 7 5 DEEP_WORK" β reward +1.49 β
β c2: "5 5 5 SLEEP" β reward -0.21 (constant belief: -0.03 β
β + sleep replay: -0.18) β
β c3: "3 6 4 ADMIN_WORK" β reward +1.10 β
β c4: "4 7 6 LEARN" β reward +1.32 β
β c5: "2 8 3 MEDITATE" β reward +0.45 (rep pen: -0.10 since β
β meditate at step 3 too) β
β c6: "3 7 7 DEEP_WORK" β reward +1.55 (belief slightly better) β
β c7: "5 5 5 EXERCISE" β reward -0.15 β
β c8: "4 6 5 FAMILY_TIME" β reward +0.92 β
β β
β group_mean = +0.81 β
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
βΌ advantages = reward - group_mean
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β ADVANTAGE (this is what GRPO actually backprops on) β
β c1: +0.68 β strongly preferred β
β c2: -1.02 β strongly discouraged (constant belief) β
β c3: +0.29 β
β c4: +0.51 β
β c5: -0.36 β
β c6: +0.74 β most preferred β
β c7: -0.96 β
β c8: +0.11 β
β β
β KEY INSIGHT: only the SPREAD matters. If all 8 had the β
β same reward, advantages would all be 0 β no gradient. β
β This is why iter 1 mode-collapsed: format_valid +1.0 for β
β every completion meant zero variance from that layer. β
βββββββββββββββ¬βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
βΌ policy loss = -E[ adv Γ log_prob(completion) ]
β + Ξ² Γ KL(policy || base_qwen)
β
βΌ backprop (only LoRA weights, ~30M params)
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β Model nudge: β
β push: "3 7 5 DEEP_WORK"-like outputs UP β
β push: "3 7 7 DEEP_WORK"-like outputs UP β
β pull: "5 5 5 *"-like outputs DOWN β
β β
β KL constraint: prevents the policy from diverging too β
β far from base Qwen (avoids gibberish drift). β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
Repeat 800-2000 times. Each step ~3-5 sec on A100.
7. The dataset is just starting positions (not labels)
DATASET (~3000 rows, generated ONCE before training)
For 200-300 episodes:
env.reset(seed=N)
for step in range(28):
record {
prompt: [system, user_for_this_state],
seed: N, ββββ replay metadata
step_index: <current step>,
action_history: <actions taken so far>,
profile_mode: "continuous",
}
env.step(rollout_policy(obs)) ββ rollout=heuristic OR random
(only matters for STATE diversity,
the agent's training generations
REPLACE these actions)
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β A row from the dataset: β
β { β
β prompt: [...full state at step 5 of episode seed=42...] β
β seed: 42, β
β step_index: 5, β
β action_history: ["deep_work", "sleep", "socialize", β
β "meditate", "deep_work"], β
β profile_mode: "continuous" β
β } β
β β
β NOTE: NO "correct action" or "label" anywhere. β
β The reward function reconstructs the env from this metadata β
β and scores whatever action the LLM picks. β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
This is fundamentally different from supervised learning:
- Supervised: (input, target_output) β model learns to mimic target
- GRPO: (prompt, replay_metadata) β model learns to maximize reward
8. Final episode grader (only fires at step 28)
_grade_episode() β runs at done=True
ββββββββββββββββ
β final_score β
β β [0, 1] β
ββββββββ¬ββββββββ
ββββββββββββββββ¬ββββββββββββββΌβββββββββββββββ¬βββββββββββββββ¬βββββββββββββββ
β β β β β β
βΌ βΌ βΌ βΌ βΌ βΌ
ββββββββββββββ ββββββββββββββ ββββββββββββ ββββββββββββββ ββββββββββββββ ββββββββββββββ
β crash_free β β progress β β connectionβ β adaptation β β efficiency β β belief β
β Γ 0.15 β β Γ 0.20 β β Γ 0.10 β β Γ 0.25 β β Γ 0.10 β β accuracy β
β β β β β β β β β β β Γ 0.20 β
ββββββββββββββ€ ββββββββββββββ€ ββββββββββββ€ ββββββββββββββ€ ββββββββββββββ€ ββββββββββββββ€
β 1 - crashesβ β final P β β final Cn β β late-half β β avg_reward β β 1 - MAE β
β /total_ck β β value β β value β β mean rewardβ β normalized β β vs true β
β β β β β β β - early β β to [0,1] β β profile β
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 β e.g. 0.80 β
β Γ0.15=0.14 β β Γ0.20=0.084β β Γ0.10=0.05β β Γ0.25=0.045β β Γ0.10=0.055β β Γ0.20=0.16 β
ββββββββββββββ ββββββββββββββ ββββββββββββ ββββββββββββββ ββββββββββββββ ββββββββββββββ
Ξ£ = 0.14 + 0.084 + 0.05 + 0.045 + 0.055 + 0.16
= 0.534 β final_score (with inference)
Heuristic / random baselines never call env.record_belief(), so the belief
component scores 0 for them β by design: the meta-RL skill is INFERENCE,
and only agents that actually try get credit on this axis.
Plus a sparse terminal reward (added to step 27's per-step reward):
terminal_bonus = (final_score - 0.5) Γ 5 β e.g. (0.534 - 0.5) Γ 5 = +0.17
This means: at step 27, agent gets last per-step reward + bonus from grader.
This is the only direct gradient signal pointing at the actual episode quality.
9. Three eval conditions (post-training)
inference_eval.py runs ALL THREE
ββββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββ¬βββββββββββββββββββββββββββββ
β discrete-3-profiles β continuous-in-distributionβ continuous-OOD β
β (3 reference profiles) β (was the agent able to β (does meta-policy β
β β learn the meta-policy?) β generalize?) β
ββββββββββββββββββββββββββββββΌβββββββββββββββββββββββββββββΌβββββββββββββββββββββββββββββ€
β env.reset(seed=N, β env.reset(seed=N) β env.reset(seed=10000+N) β
β profile="introvert_ β β samples from training β β samples from a region β
β morning") β distribution β never seen in trainingβ
β β β β
β β 3 hardcoded profiles β β ~10 sampled profiles β β ~10 sampled profiles β
β (from PROFILES list) β from seeds 100..110 β from seeds 10000..10010 β
β β β β
β Each strategy plays each β Each strategy plays each β Each strategy plays each β
β profile 5x = 15 episodes β seed 1x = 10 episodes β seed 1x = 10 episodes β
β β β β
β Strategies tested: β Strategies tested: β Strategies tested: β
β β’ random β β’ random β β’ random β
β β’ heuristic β β’ heuristic β β’ heuristic β
β β’ model (trained Qwen) β β’ model (trained Qwen) β β’ model (trained Qwen) β
ββββββββββββββββββββββββββββββ΄βββββββββββββββββββββββββββββ΄βββββββββββββββββββββββββββββ
THE KEY METRIC: trained model's score on continuous-OOD vs heuristic baseline.
Heuristic baseline (profile-blind hand rules): score 0.580 on OOD.
Trained meta-RL agent's target: > 0.580 on OOD.
If the trained agent beats the heuristic on OOD (profiles never seen in
training), that's direct proof it learned the SKILL of profile inference,
not just memorized training profiles.
10. Sim β Real mapping (why this POC matters)
The whole point of training in this env is that the SKILL transfers. The mapping from simulated signals to production signals is direct: same information, different source.
ββββββββββββββββββββββββ SIMULATION (training) ββββββββ βββββ REAL PRODUCT (deployment) ββββββββ
β β β β
β meter: Vitality = 0.62 ββββΌββΊ HRV from watch + sleep score β
β meter: Cognition = 0.51 ββββΌββΊ focus app metrics + screen time β
β meter: Progress = 0.24 ββββΌββΊ calendar density + task completionβ
β meter: Serenity = 0.71 ββββΌββΊ HRV variability + voice sentiment β
β meter: Connection = 0.38 ββββΌββΊ call/message freq + calendar β
β β β β
β hidden profile (sampled per episode) ββββΌββΊ the real user's actual personalityβ
β ββ never visible to agent β β ββ also never visible β
β β β β
β agent emits: "3 7 5 DEEP_WORK" β β β
β ββ "3 7 5" = belief about user (hidden internal) ββββΌββΊ agent's internal user model β
β ββ "DEEP_WORK" = action choice ββββΌββΊ recommendation to user β
β β β ("I suggest a focus block now") β
β β β β
β per-step env_reward ββββΌββΊ meter trend + accept/ignore tap β
β ββ "did meters improve under profile weights?" β β ββ "did user accept and benefit?"β
β β β β
β episode = 1 week (28 steps) ββββΌββΊ rolling weekly window β
β β β β
β 3 eval conditions β β USER ARRIVES (cold start) β
β ββ discrete-3 (memorization check) β β ββ Agent has zero info about them β
β ββ continuous-in-dist (training-distrib) β β ββ ~5-10 interactions to converge β
β ββ continuous-OOD βββ KEY βββΊ ββββΌββΊ on a confident belief vector β
β agent must infer profile NEVER seen in β β ββ Acts on belief, learns from β
β training. THIS is the production scenario. β β tap responses β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββ βββββββββββββββββββββββββββββββββββββββββ
SUCCESS METRIC IN BOTH WORLDS: how fast does the agent personalize?
Sim: belief_accuracy curve over a single episode (does it climb?)
adaptation_score (late-half mean reward > early-half mean reward?)
Real: how many interactions until user's accept-rate stabilizes high?
How quickly does the agent stop suggesting things the user always ignores?
Both reduce to the same skill: form a model of this person from limited
observation, then act on it.
Design choices that explicitly preserve the passive-only constraint:
- β No "ASK USER" action β agent never quizzes
- β No profile feature in observation at eval time β agent must infer
- β No explicit "what's your goal?" prompt β only sensor-equivalent meters
- β Belief output is INTERNAL (visible in POC only for training reward)
- β Reward formula uses only quantities computable from passive signals
If the agent learns to do well on continuous-OOD (profiles never seen in training), it has acquired the skill of "figure out a new person from observation alone" β exactly the skill the production assistant needs when meeting a real user for the first time.
11. Spend & timing (concrete)
HARDWARE: A100 large (80GB) on HF Jobs at $2.50/hr ($0.0417/min)
FAST_MODE (200-800 steps):
dataset gen: ~2 min
model load: ~3 min
training: ~10-25 min (depends on steps)
eval: ~3 min
plot + upload: ~2 min
ββββββββββββββββββββββββββββ
total: 20-35 min ($0.80-1.50 per iter)
FULL RUN (2000 steps):
dataset gen: ~3 min
model load: ~3 min
training: ~60-90 min
eval: ~3 min
plot + upload: ~2 min
ββββββββββββββββββββββββββββ
total: 70-100 min ($3-4)
Iter 1 (200 steps): $0.50 β mode collapse (single action)
Iter 2 (400 steps): $1.50 β mode collapse (2-cycle)
Iter 3 (800 steps): $5 β³ in flight (control)
Iter 4 (800 steps): $5 β³ in flight (with full fixes)
Final (2000 steps): $4 β³ pending iter 3+4 results
ββββββ
~$16 of $30 budget