Four intuitive high-quality Harrier mixtures with explicit two-phase cXXqY weights

MIXES.md

Four intuitive high-quality Harrier mixtures

These are four new 18.75T-token, two-phase priors for d768 seed 0. Phase 0 is 15T tokens (80%); phase 1 starts at step 8448 and is 3.75T tokens (20%). The quality assumption is monotonic: q4 is best, followed by q3, q2, q1, and q0.

The recipes are intentionally prior-driven. Earlier experiments inform the cumulative epoch cap: every cXXqY source stays at or below four epochs across both phases. Use the CSV weights verbatim; its 400 weight columns are the complete phase-0 and phase-1 mixtures.

Recipes

Family shares are ordered code_agent / science_math / knowledge / low_value / general. Quality shares are ordered q0 / q1 / q2 / q3 / q4.

I1: Broad quality staircase

Run: harrier-intuitive-i1-broad-quality-staircase-d768-seed0

Replacing low-quality data with higher-quality data improves aggregate performance while proportional domain coverage preserves breadth.

Every domain keeps its exact available-token share in both phases; only the quality distribution changes.

• Phase 0 family: 25.0% / 8.7% / 19.8% / 11.8% / 34.7%
• Phase 0 quality: 0.0% / 5.0% / 15.0% / 35.0% / 45.0%
• Phase 1 family: 25.0% / 8.7% / 19.8% / 11.8% / 34.7%
• Phase 1 quality: 0.0% / 0.0% / 10.0% / 30.0% / 60.0%
• Maximum cumulative cell epochs: 4.000000

I2: High-quality AA portfolio

Run: harrier-intuitive-i2-aa-quality-portfolio-d768-seed0

A high-quality portfolio tilted toward code, agents, science, math, and knowledge improves broad intelligence-index performance.

• Phase 0 family: 30.0% / 18.0% / 25.0% / 5.0% / 22.0%
• Phase 0 quality: 0.0% / 3.0% / 12.0% / 35.0% / 50.0%
• Phase 1 family: 36.0% / 24.0% / 24.0% / 2.0% / 14.0%
• Phase 1 quality: 0.0% / 0.0% / 5.0% / 30.0% / 65.0%
• Maximum cumulative cell epochs: 4.000000

I3: High-quality textbook and reference

Run: harrier-intuitive-i3-textbook-reference-d768-seed0

High-quality structured knowledge and scientific material transfers broadly across reasoning and knowledge benchmarks.

• Phase 0 family: 15.0% / 20.0% / 40.0% / 5.0% / 20.0%
• Phase 0 quality: 0.0% / 3.0% / 12.0% / 40.0% / 45.0%
• Phase 1 family: 15.0% / 25.0% / 45.0% / 2.0% / 13.0%
• Phase 1 quality: 0.0% / 0.0% / 5.0% / 35.0% / 60.0%
• Maximum cumulative cell epochs: 4.000000

I4: Conservative high-quality anti-junk

Run: harrier-intuitive-i4-conservative-quality-d768-seed0

Suppressing the lowest-value material and applying a moderate technical tilt captures quality gains while retaining broad general coverage.

• Phase 0 family: 27.0% / 10.0% / 22.0% / 6.0% / 35.0%
• Phase 0 quality: 2.0% / 8.0% / 22.0% / 38.0% / 30.0%
• Phase 1 family: 30.0% / 14.0% / 26.0% / 3.0% / 27.0%
• Phase 1 quality: 0.0% / 3.0% / 12.0% / 35.0% / 50.0%
• Maximum cumulative cell epochs: 2.792657

Domain families

The five families are disjoint and cover c00-c39 exactly.

• code_agent: c13 (Command-line agent transcripts), c14 (Software development and web code), c22 (Application and web development), c28 (Performance logs and low-level code), c32 (Software infrastructure and security)
• science_math: c15 (Clinical treatment and dentistry), c30 (Natural sciences research), c39 (Mathematics problems and proofs)
• knowledge: c00 (Geopolitics and world news), c02 (Arts, history, and culture), c05 (Government and environment), c06 (Public policy), c09 (Study skills and academic writing), c20 (School curricula and teacher training), c27 (General knowledge and trivia), c29 (Courses and tutoring), c37 (Law, courts, and regulation), c38 (History, literature, and heritage)
• low_value: c07 (Games and gambling), c17 (SEO and online business), c25 (Whimsical and fan fiction), c33 (Racing, sports, and entertainment records), c36 (Gambling and financial services)
• general: c01 (Consumer health and supplements), c03 (Finance, insurance, and markets), c04 (Consumer tech and cybersecurity), c08 (Schools and children's programs), c10 (Personal wellbeing and conduct), c11 (Food and recipes), c12 (Travel and destinations), c16 (Sports and competition), c18 (Business software and strategy), c19 (Home maintenance and trades), c21 (Christian faith and devotion), c23 (Consumer devices and vehicles), c24 (Emotions, relationships, and wellbeing), c26 (Council and board meeting records), c31 (Fiction, film, and television), c34 (Pets, gardening, and pest control), c35 (Crafts, design, and home decor)

Allocation and validation

I1 uses a 40x5 domain-quality transportation problem so every domain share remains exactly proportional. I2-I4 use 5x5 family-quality transportation problems and distribute each family-quality allocation among its cells in proportion to available tokens. Each solution stays as close as possible to the available-token distribution subject to its exact marginals and the four-epoch cumulative cap.

validation.json records the requested and achieved marginals, phase sums, maximum cumulative epoch count, and any cells that reach the cap. All phase weight sums equal 1 within 1e-10, all requested marginals match within 1e-8, and all cumulative cell epoch counts are at most 4.

two_phase_quality_prior_weights.csv

validation.json

[
  {
    "run_name": "harrier-intuitive-i1-broad-quality-staircase-d768-seed0",
    "phase_tokens": [
      15000000000000,
      3750000000000
    ],
    "phase_weight_sums": [
      1.0,
      1.0
    ],
    "requested_family_shares": [
      [
        0.25014637965297604,
        0.08729950477967693,
        0.19773251723515395,
        0.11792816321461089,
        0.3468934351175822
      ],
      [
        0.25014637965297604,
        0.08729950477967693,
        0.19773251723515395,
        0.11792816321461089,
        0.3468934351175822
      ]
    ],
    "achieved_family_shares": [
      [
        0.25014637965297654,
        0.08729950477967738,
        0.197732517235153,
        0.11792816321461091,
        0.3468934351175821
      ],
      [
        0.25014637965298353,
        0.08729950477967743,
        0.19773251723515572,
        0.11792816321460808,
        0.34689343511757526
      ]
    ],
    "requested_quality_shares": [
      [
        0.0,
        0.05,
        0.15,
        0.35,
        0.45
      ],
      [
        0.0,
        0.0,
        0.1,
        0.3,
        0.6
      ]
    ],
    "achieved_quality_shares": [
      [
        0.0,
        0.049999999999999996,
        0.14999999999999997,
        0.35,
        0.45000000000000007
      ],
      [
        0.0,
        0.0,
        0.1,
        0.3,
        0.6000000000000001
      ]
    ],
    "maximum_cumulative_cell_epochs": 4.0000000000102265,
    "cells_at_epoch_cap": [
      "c00q4",
      "c03q4",
      "c04q4",
      "c07q4",
      "c09q4",
      "c10q4",
      "c11q4",
      "c12q4",
      "c14q4",
      "c16q4",
      "c17q4",
      "c18q4",
      "c19q4",
      "c20q4",
      "c21q4",
      "c22q4",
      "c23q4",
      "c24q4",
      "c25q4",
      "c27q4",
      "c28q4",
      "c29q4",
      "c32q4",
      "c33q4",
      "c35q4",
      "c36q4",
      "c38q4"
    ],
    "requested_domain_shares": [
      [
        0.034142146381366804,
        0.03706080239325818,
        0.010479839450363947,
        0.034684633643826936,
        0.008471697509869591,
        0.023721213353246406,
        0.010156644633474257,
        0.012947898908722222,
        0.006668206357970776,
        0.007788941258051893,
        0.011239011272034152,
        0.012471521817070955,
        0.0177954890942776,
        0.026326934787244557,
        0.0757098399035683,
        0.0133478213887454,
        0.015479007765341871,
        0.025577146651036217,
        0.010061179078726196,
        0.011500085816304503,
        0.02088482537751864,
        0.020980095971293123,
        0.005033763639092895,
        0.02697366397681158,
        0.021975225320519192,
        0.039623315077679806,
        0.07772958838082165,
        0.0026303102666201967,
        0.07215346872646192,
        0.02542060789914409,
        0.05031392633794227,
        0.00682021272051456,
        0.07092237259660836,
        0.032436558854774596,
        0.006223130860050946,
        0.020759883138890402,
        0.007343243722398042,
        0.01959376325244796,
        0.04291422536291974,
        0.02363775705298925
      ],
      [
        0.034142146381366804,
        0.03706080239325818,
        0.010479839450363947,
        0.034684633643826936,
        0.008471697509869591,
        0.023721213353246406,
        0.010156644633474257,
        0.012947898908722222,
        0.006668206357970776,
        0.007788941258051893,
        0.011239011272034152,
        0.012471521817070955,
        0.0177954890942776,
        0.026326934787244557,
        0.0757098399035683,
        0.0133478213887454,
        0.015479007765341871,
        0.025577146651036217,
        0.010061179078726196,
        0.011500085816304503,
        0.02088482537751864,
        0.020980095971293123,
        0.005033763639092895,
        0.02697366397681158,
        0.021975225320519192,
        0.039623315077679806,
        0.07772958838082165,
        0.0026303102666201967,
        0.07215346872646192,
        0.02542060789914409,
        0.05031392633794227,
        0.00682021272051456,
        0.07092237259660836,
        0.032436558854774596,
        0.006223130860050946,
        0.020759883138890402,
        0.007343243722398042,
        0.01959376325244796,
        0.04291422536291974,
        0.02363775705298925
      ]
    ],
    "achieved_domain_shares": [
      [
        0.034142146381366305,
        0.03706080239325804,
        0.010479839450363921,
        0.03468463364382709,
        0.008471697509869744,
        0.02372121335324622,
        0.010156644633474038,
        0.012947898908722423,
        0.006668206357970611,
        0.007788941258051744,
        0.011239011272034161,
        0.012471521817070577,
        0.01779548909427764,
        0.026326934787245168,
        0.07570983990356846,
        0.013347821388745906,
        0.015479007765341782,
        0.02557714665103645,
        0.010061179078726099,
        0.01150008581630469,
        0.02088482537751868,
        0.0209800959712931,
        0.005033763639092756,
        0.026973663976811724,
        0.021975225320519307,
        0.03962331507767956,
        0.07772958838082153,
        0.0026303102666201256,
        0.07215346872646158,
        0.02542060789914434,
        0.050313926337942454,
        0.006820212720514702,
        0.07092237259660861,
        0.0324365588547744,
        0.0062231308600507575,
        0.020759883138890548,
        0.007343243722398086,
        0.019593763252447556,
        0.0429142253629201,
        0.023637757052989015
      ],
      [
        0.034142146381365555,
        0.03706080239325743,
        0.010479839450366632,
        0.03468463364382646,
        0.00847169750987016,
        0.023721213353246673,
        0.010156644633475398,
        0.012947898908721937,
        0.006668206357972975,
        0.007788941258049232,
        0.011239011272033778,
        0.012471521817069052,
        0.017795489094277278,
        0.026326934787246334,
        0.07570983990356744,
        0.013347821388746491,
        0.015479007765339423,
        0.025577146651036598,
        0.010061179078725273,
        0.011500085816303851,
        0.020884825377517645,
        0.020980095971292717,
        0.005033763639099841,
        0.026973663976810537,
        0.021975225320518377,
        0.03962331507767939,
        0.07772958838082059,
        0.0026303102666231475,
        0.07215346872646133,
        0.025420607899143854,
        0.05031392633794233,
        0.0068202127205161615,
        0.07092237259660855,
        0.03243655885477402,
        0.006223130860052056,
        0.020759883138889136,
        0.007343243722396128,
        0.019593763252447334,
        0.04291422536292027,
        0.023637757052988612
      ]
    ]
  },
  {
    "run_name": "harrier-intuitive-i2-aa-quality-portfolio-d768-seed0",
    "phase_tokens": [
      15000000000000,
      3750000000000
    ],
    "phase_weight_sums": [
      1.0,
      0.9999999999999998
    ],
    "requested_family_shares": [
      [
        0.3,
        0.18,
        0.25,
        0.05,
        0.22
      ],
      [
        0.36,
        0.24,
        0.24,
        0.02,
        0.14
      ]
    ],
    "achieved_family_shares": [
      [
        0.30000000000000004,
        0.17999999999999997,
        0.24999999999999994,
        0.049999999999999954,
        0.22000000000000006
      ],
      [
        0.35999999999999976,
        0.24000000000000007,
        0.24000000000000002,
        0.02000000000000008,
        0.13999999999999999
      ]
    ],
    "requested_quality_shares": [
      [
        0.0,
        0.03,
        0.12,
        0.35,
        0.5
      ],
      [
        0.0,
        0.0,
        0.05,
        0.3,
        0.65
      ]
    ],
    "achieved_quality_shares": [
      [
        0.0,
        0.029999999999999995,
        0.11999999999999998,
        0.35,
        0.5
      ],
      [
        0.0,
        0.0,
        0.05,
        0.30000000000000004,
        0.6499999999999999
      ]
    ],
    "maximum_cumulative_cell_epochs": 4.000000000000001,
    "cells_at_epoch_cap": [
      "c00q4",
      "c02q4",
      "c05q4",
      "c06q4",
      "c09q4",
      "c13q4",
      "c14q4",
      "c15q4",
      "c20q4",
      "c22q4",
      "c27q4",
      "c28q4",
      "c29q4",
      "c30q4",
      "c32q4",
      "c37q4",
      "c38q4",
      "c39q4"
    ]
  },
  {
    "run_name": "harrier-intuitive-i3-textbook-reference-d768-seed0",
    "phase_tokens": [
      15000000000000,
      3750000000000
    ],
    "phase_weight_sums": [
      0.9999999999999999,
      1.0
    ],
    "requested_family_shares": [
      [
        0.15,
        0.2,
        0.4,
        0.05,
        0.2
      ],
      [
        0.15,
        0.25,
        0.45,
        0.02,
        0.13
      ]
    ],
    "achieved_family_shares": [
      [
        0.15,
        0.2000000000000001,
        0.39999999999999986,
        0.05000000000000003,
        0.1999999999999999
      ],
      [
        0.1500000000000001,
        0.25000000000000033,
        0.4499999999999994,
        0.020000000000000302,
        0.12999999999999998
      ]
    ],
    "requested_quality_shares": [
      [
        0.0,
        0.03,
        0.12,
        0.4,
        0.45
      ],
      [
        0.0,
        0.0,
        0.05,
        0.35,
        0.6
      ]
    ],
    "achieved_quality_shares": [
      [
        0.0,
        0.029999999999999995,
        0.12,
        0.39999999999999997,
        0.44999999999999996
      ],
      [
        0.0,
        0.0,
        0.05,
        0.35,
        0.6
      ]
    ],
    "maximum_cumulative_cell_epochs": 4.000000000000004,
    "cells_at_epoch_cap": [
      "c00q4",
      "c02q4",
      "c05q4",
      "c06q4",
      "c09q4",
      "c15q4",
      "c20q4",
      "c27q4",
      "c29q4",
      "c30q4",
      "c37q4",
      "c38q4",
      "c39q4"
    ]
  },
  {
    "run_name": "harrier-intuitive-i4-conservative-quality-d768-seed0",
    "phase_tokens": [
      15000000000000,
      3750000000000
    ],
    "phase_weight_sums": [
      1.0,
      1.0
    ],
    "requested_family_shares": [
      [
        0.27,
        0.1,
        0.22,
        0.06,
        0.35
      ],
      [
        0.3,
        0.14,
        0.26,
        0.03,
        0.27
      ]
    ],
    "achieved_family_shares": [
      [
        0.27,
        0.09999999999999999,
        0.22,
        0.05999999999999999,
        0.35000000000000003
      ],
      [
        0.30000000000000004,
        0.14000000000000024,
        0.25999999999999984,
        0.02999999999999989,
        0.2700000000000001
      ]
    ],
    "requested_quality_shares": [
      [
        0.02,
        0.08,
        0.22,
        0.38,
        0.3
      ],
      [
        0.0,
        0.03,
        0.12,
        0.35,
        0.5
      ]
    ],
    "achieved_quality_shares": [
      [
        0.02,
        0.08000000000000002,
        0.21999999999999997,
        0.38,
        0.30000000000000004
      ],
      [
        0.0,
        0.03,
        0.11999999999999997,
        0.35000000000000003,
        0.5000000000000001
      ]
    ],
    "maximum_cumulative_cell_epochs": 2.792657090587314,
    "cells_at_epoch_cap": []
  }
]