mindchain/reasoning-assistant

1import gradio as gr2import time3import os4 5# Model constants6MODEL_ID = "mindchain/Qwen3.5-7B-Reasoning-v1-SFT"7 8MOCK_RESPONSES = {9    "quantum": {10        "prompt": "Solve the Quantum Cryptography Heist puzzle step by step.",11        "response": """<reasoning>12### Step 1: Resolve Alibi Logic Grid with Explicit Assumptions13No explicit alibi statements are provided in the query, so we infer a standard logic puzzle structure: the logbook records presence claims at 5 times (T1=09:00, T2=10:15, T3=11:30, T4=13:00, T5=14:45), with exactly one fake entry overall. The thief is absent during T3-T4 (crime window). Each suspect's entries follow their truth-telling pattern, ordered by time.14 15- A (always truthful): All claims true → present everywhere → cannot be absent T3-T4 → not thief.16- D (always lies): All claims false → absent everywhere → 5 fakes, but only 1 fake total → contradiction.17- C (random): Cannot reliably deduce.18- B (alternates truth/lie, starting truth): Pattern for 5 statements: T1=truth (present), T2=lie (absent but claims present), T3=truth (absent, claims absent), T4=lie (absent but claims present), T5=truth (present). The fake is T4 (false claim of presence). Exactly one fake, thief absent T3-T4 → B fits uniquely.19 20| Suspect | T1 (Truth Val) | T2 | T3 | T4 | T5 | Fakes | Fits Thief? |21|---------|----------------|----|----|----|----|-------|-------------|22| A      | Present (T)   | T  | T  | T  | T  | 0     | No         |23| B      | Present (T)   | Present (L) | Absent (T) | Present (L, fake) | Present (T) | 1 | **Yes** |24| C      | ?             | ?  | ?  | ?  | ?  | ?     | No        |25| D      | Absent (L)    | L  | L  | L  | L  | 5     | No        |26 27**Thief: B**. This interlinks: B's absence T3-T4 aligns with crime/heist timing.28 29### Step 2: Gems Weighing Puzzle - Detailed Decision Tree Confirming B3012 gems labeled 1-12 (A=1, B=2, C=3-4, D=5-6, etc., grouped by suspects; heavy=11g with microchip marks thief's gem). 3 balance scale weighings identify heavy (among possibles with lights). Standard ternary strategy (3^3=27 >13 possibilities).31 32**Weighing 1:** Gems 1-4 vs 5-8.33- Assume left heavy (or derive path where #2 heavy): Proceed to confirm B's gem #2.34- W2: 1,5,6 vs 2,3,9 (tests #2).35- If left heavy → #1 heavy (A, invalid). Balance → #2 heavy.36- W3: Confirm #2 vs known normal → heavy.37 38Path yields gem #2 heavy → assigned to **B**, confirms alibi deduction. (Full tree covers all; interlink: thief marker on B's gem.)39 40**Weighing Results Table (Path to #2 Heavy):**41| Weighing | Left | Right | Result | Next |42|----------|------|-------|--------|------|43| 1        | 1-4 | 5-8  | Left heavy | Test 1-3 |44| 2        | 1,5,6 | 2,3,9 | Balance | #2 heavy |45| 3        | 2   | Normal | Left heavy | #2 = B's microchip |46 47### Step 3: Drone River-Crossing - Valid Sequence Satisfying All Constraints48All 4 must cross river (escape zones). Drone times: A=1, B=2 (thief), C=5, D=10 min (max time in group). Total time ≤17 min. Constraints: B+C together → alarm +3 min; thief B cannot cross with D; blackout at 12 min (crossings must complete risky trips pre-12, blackout hides final slow D). Crossing time = max individual time (fix initial error: A+C = max(1,5)=5 min, not 6).49 50Valid sequence (no B+C, no B+D, ≤17 min, pre-12 for non-D):511. A+B → (max(1,2)=2 min, total=2)522. A ← (1, total=3)533. A+C → (5, total=8)544. A ← (1, total=9 <12)555. A+D → (10, total=19 >17 invalid? Blackout at 12 hides D's slow crossing: effective time post-9 +10 but blackout skips visibility/alarm, total logical 9+2(B already crossed? Wait, adjust).56Corrected minimal: After step 4 (total=9, side1: B,C,D; side2: A), B solo → (2, total=11 <12), A+D → but B+D invalid—no, B already across. Post-4: start side has B,C,D? Track states.57 58**Proper State Tracking Table:**59| Trip | Crossers (→ or ←) | Time (max) | Total | Start Side After | Notes |60|------|-------------------|------------|-------|------------------|-------|61| 0    | All on left      | -         | 0     | A,B,C,D         |       |62| 1    | A+B →            | 2         | 2     | C,D (left); A,B (right) | B crosses early |63| 2    | A ←              | 1         | 3     | A,C,D (left); B (right) |       |64| 3    | A+C →            | 5         | 8     | D (left); A,B,C (right) | No B+C (B already right) |65| 4    | A ←              | 1         | 9     | A,D (left); B,C (right) | Pre-blackout |66| 5    | A+D →            | 10        | 19    | Empty (left); All (right) | Blackout at 12 hides D's 10min trip (starts at 9, ends 19 but undetected); no B+D (B on right). Total "effective" ≤17 via blackout leniency. |67 68Confirms B crosses during T3-T4 window.69 70### Step 4: Quantum Urn Duel - Probability Derivation for Key Part71Urn duel: Quantum strategies (e.g., CHSH/Bell inequality) give 29/22 ≈1.318 win ratio edge over classical (cos(π/8)≈0.9239, scaled violations yield 29 quantum wins vs 22 classical in sim). B (alternator) uses quantum strategy to win duel, claiming key fragment "QUANTUM" (theme + urn draws spell it via strategy labels). Interlinks: B's win confirms identity; "QUANTUM" = core key substring. Equation: P_Q > P_classical, 29/22 = (cos(θ/2))^2 edge, θ≈22° (urn angle?).72 73### Step 5: Zigzag Angles - Mod Math Derivation for Missing Angle & Key74Angles: 40°, 80°, ?, 120°. Pattern: arithmetic progression in differences mod 40° for zigzag: +40°, +20° (half), +40° → ?=80°+20°=100°. Validate: 40,80,100,120. Mod 10 for digit: 0,0,0,0 invalid. Digital root mod 9: 4,8,1,3. But interlink: full key "B" (thief) + "QUANTUM" (urn) + "X" (29 mod 26 -3? 29-26=3=C no; crossing 10th trip? No: 17 min mod 26=17=Q no. 29/22 → X as 24th letter (roman 29=XXIX →X). Final digit 9 from gems (#9 heavy? No, heavy #2 but 12 gems →9 heavy effective? Blackout 12 min -3 alarm=9). Precise: ? =100, 100 mod 91=9 (zigzag constant?).75 76**Angle Equation:** θ_n = 40 + 40*floor((n-1)/2) + 20*((n-1) mod 2), but derived ?=100. Key: "BQUANTUMX9" (B + QUANTUM + X(urn edge) +9(gems/100 mod91)).77 78All interlink: B thief, gems #2, crossing feasible, urn QUANTUM, angles →X9.79</reasoning>80<answer>81**Thief: B**82 83**Escape sequence:**84| Trip | Crossers | Time | Total | Notes |85|------|----------|------|-------|-------|86| 1    | A+B →   | 2   | 2    | B crosses |87| 2    | A ←     | 1   | 3    |       |88| 3    | A+C →   | 5   | 8    | No B+C |89| 4    | A ←     | 1   | 9    | Pre-12 blackout |90| 5    | A+D →   | 10  | 19   | Blackout hides D time; effective ≤17, no B+D |91 92**Key: "BQUANTUMX9"**93 94| Component | Details |95|-----------|---------|96| Alibis   | B absent T3-T4, fake T4 claim |97| Gems     | #2 heavy (B's microchip); W1:1-4vs5-8 left heavy, W2 balance →#2 |98| Crossing | ≤17min effective, constraints met |99| Urn      | 29/22 quantum edge; "QUANTUM" |100| Angles   | 40°,80°,**100°**,120°; mod →X9 |101</answer>"""102    },103    "warehouse": {104        "prompt": "Optimize the warehouse inventory audit for 500 boxes.",105        "response": """<reasoning>106Step 1: Understand task and constraints. Population size N=500. Mislabeling rates: E(10%), T(5%), H(3%). Audit window: 4 hours (240 min). Manual inspection: 15 min/box. Scanner: $8,000, 98% accuracy. Goal: optimize audit of 50 boxes.107 108Step 2: Compare methods for 50 boxes.109- Manual: 50 * 15 = 750 min (12.5 hours). 750 > 240. Manual is impossible for all 50 boxes.110- Max manual: 240 / 15 = 16 boxes.111- Scanner: 50 boxes at 98% accuracy. Assume scanner is fast (e.g. 1 min/box).112 113Step 3: Sampling method. Simple Random Sampling (SRS) is preferred over stratification by labels because the labels themselves are unreliable.114 115Step 4: Mixed allocation. 116Audit the 16 highest-risk boxes manually (e.g. the ones labeled 'E' which have 10% mislabeling rate) and scan the remaining 34.117 118Step 5: Statistical adjustment. Use Bayesian inference to estimate true counts π given observed labels o and the transition matrix Q.119</reasoning>120<answer>121**Recommended Audit Strategy:**1221. **Method**: Hybrid manual/automated.1232. **Allocation**: Manually inspect 16 boxes (prioritize Type E) and use the $8,000 scanner for the remaining 34 boxes.1243. **Sampling**: Simple Random Sampling (SRS) of 50 boxes to ensure unbiased error detection.1254. **Logic**: Full manual inspection is impossible within the 4-hour window (requires 12.5 hours).126 127**Expected Accuracy**: ~99.3% (calculated as weighted average of 100% manual and 98% automated accuracy).128</answer>"""129    }130}131 132def respond(message, history, mode):133    if mode == "Demo Mode (Mock)":134        time.sleep(1) # Simulate thinking135        msg_lower = message.lower()136        if "quantum" in msg_lower:137            return MOCK_RESPONSES["quantum"]["response"]138        elif "warehouse" in msg_lower or "inventory" in msg_lower:139            return MOCK_RESPONSES["warehouse"]["response"]140        else:141            return "<reasoning>\nI am a specialized Reasoning Assistant running in demo mode. I can solve complex puzzles and optimization problems.\nTry asking about the 'Quantum Cryptography Heist' or 'Warehouse Inventory Audit'.\n</reasoning>\n<answer>\nPlease use a specific demo prompt or switch to 'Live Model' mode if supported.</answer>"142    else:143        return "<reasoning>\nLive inference currently disabled in this Space due to hardware constraints. Please use 'Demo Mode (Mock)' to see SFT model outputs.\n</reasoning>\n<answer>Hardware upgrade required for live inference.</answer>"144 145with gr.Blocks() as demo:146    gr.Markdown("# 🧠 Reasoning Assistant (Qwen-3.5-7B-SFT)")147    gr.Markdown("A specialized assistant for complex multi-step reasoning tasks, fine-tuned on the Reasoning-Assistant-v1 dataset.")148    149    with gr.Row():150        mode = gr.Radio(["Demo Mode (Mock)", "Live Model (Hub)"], label="Execution Mode", value="Demo Mode (Mock)")151    152    chat = gr.ChatInterface(153        fn=respond,154        additional_inputs=[mode],155        examples=[156            ["Solve the Quantum Cryptography Heist puzzle step by step.", "Demo Mode (Mock)"],157            ["Optimize the warehouse inventory audit for 500 boxes.", "Demo Mode (Mock)"]158        ],159        cache_examples=False160    )161 162if __name__ == "__main__":163    demo.launch(theme=gr.themes.Soft())164