Affine-sharp4 / babyai_validation_layers.md
sharp1112's picture
Upload folder using huggingface_hub
ebec87c verified

BABYAI Environment: Validation Layer Analysis

Overview

BABYAI is one of the 8 validation environments. This document shows which validation scoring layers include BABYAI and how it contributes to the overall scoring system.

BABYAI Environment Details

  • Environment Name: agentgym:babyai
  • Type: AgentGym environment
  • Dataset Size: 500 tasks
  • Daily Sampling Rate: 120/day (fast environment)
  • Task Type: Grid-world navigation and instruction following
  • Max Rounds: 10
  • Timeout: 1200 seconds

Validation Layers That Include BABYAI

BABYAI appears in validation layers 3-8 as part of various environment combinations. Here's the breakdown:

Layer 3 (3-environment combinations)

Total subsets with BABYAI: C(7,2) = 21 combinations

BABYAI appears in 21 out of 56 total Layer 3 subsets. Examples:

  • {BABYAI, WEBSHOP, ALFWORLD}
  • {BABYAI, SCIWORLD, TEXTCRAFT}
  • {BABYAI, SAT, DED}
  • {BABYAI, ABD, WEBSHOP}
  • ... (17 more combinations)

Weight per subset: 8.0 / 56 = 0.143 Total potential weight from BABYAI Layer 3 subsets: 21 × 0.143 = 3.003

Layer 4 (4-environment combinations)

Total subsets with BABYAI: C(7,3) = 35 combinations

BABYAI appears in 35 out of 70 total Layer 4 subsets. Examples:

  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD}
  • {BABYAI, TEXTCRAFT, SAT, DED}
  • {BABYAI, ABD, WEBSHOP, ALFWORLD}
  • ... (32 more combinations)

Weight per subset: 16.0 / 70 = 0.229 Total potential weight from BABYAI Layer 4 subsets: 35 × 0.229 = 8.015

Layer 5 (5-environment combinations)

Total subsets with BABYAI: C(7,4) = 35 combinations

BABYAI appears in 35 out of 56 total Layer 5 subsets. Examples:

  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT}
  • {BABYAI, SAT, DED, ABD, WEBSHOP}
  • ... (33 more combinations)

Weight per subset: 32.0 / 56 = 0.571 Total potential weight from BABYAI Layer 5 subsets: 35 × 0.571 = 19.985

Layer 6 (6-environment combinations)

Total subsets with BABYAI: C(7,5) = 21 combinations

BABYAI appears in 21 out of 28 total Layer 6 subsets. Examples:

  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT, SAT}
  • {BABYAI, DED, ABD, WEBSHOP, ALFWORLD, SCIWORLD}
  • ... (19 more combinations)

Weight per subset: 64.0 / 28 = 2.286 Total potential weight from BABYAI Layer 6 subsets: 21 × 2.286 = 48.006

Layer 7 (7-environment combinations)

Total subsets with BABYAI: C(7,6) = 7 combinations

BABYAI appears in 7 out of 8 total Layer 7 subsets. Examples:

  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT, SAT, DED}
  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT, SAT, ABD}
  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT, DED, ABD}
  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, SAT, DED, ABD}
  • {BABYAI, WEBSHOP, ALFWORLD, TEXTCRAFT, SAT, DED, ABD}
  • {BABYAI, WEBSHOP, SCIWORLD, TEXTCRAFT, SAT, DED, ABD}
  • {BABYAI, ALFWORLD, SCIWORLD, TEXTCRAFT, SAT, DED, ABD}

Weight per subset: 128.0 / 8 = 16.0 Total potential weight from BABYAI Layer 7 subsets: 7 × 16.0 = 112.0

Layer 8 (All 8 environments)

Total subsets with BABYAI: C(7,7) = 1 combination

BABYAI appears in the single Layer 8 subset:

  • {BABYAI, WEBSHOP, ALFWORLD, SCIWORLD, TEXTCRAFT, SAT, DED, ABD}

Weight per subset: 256.0 / 1 = 256.0 Total potential weight from BABYAI Layer 8 subset: 1 × 256.0 = 256.0

Summary Table

Layer Total Subsets Subsets with BABYAI Weight per Subset Total BABYAI Weight
3 56 21 0.143 3.003
4 70 35 0.229 8.015
5 56 35 0.571 19.985
6 28 21 2.286 48.006
7 8 7 16.0 112.0
8 1 1 256.0 256.0
Total 219 120 - 447.009

Key Insights

  1. BABYAI Coverage: BABYAI appears in 120 out of 219 total evaluated subsets (54.8% of all subsets)

  2. Exponential Importance: As layers increase, BABYAI's potential contribution grows exponentially:

    • Layer 3: 3.003 total weight
    • Layer 8: 256.0 total weight (85× more!)
  3. Comprehensive Performance Matters: To maximize BABYAI-related rewards, a model must:

    • Perform well on BABYAI alone (but this isn't evaluated in layers 1-2)
    • Perform well on BABYAI + 2 other environments (Layer 3)
    • Perform well on BABYAI + 3 other environments (Layer 4)
    • ...
    • Perform well on ALL 8 environments including BABYAI (Layer 8) - highest reward!
  4. Layer 8 Dominance: The single Layer 8 subset (all environments) contributes 256.0 weight, which is more than all other BABYAI-related subsets combined (191.009).

Relationship to 36 Transformer Layers

The 36 transformer layers in the model architecture are not directly mapped to BABYAI. Instead:

  1. All 36 layers work together to process BABYAI tasks
  2. BABYAI performance is evaluated across all 8 environments
  3. Validation scoring layers (3-8) reward models that perform well on BABYAI in combination with other environments

However, based on the codebase documentation (BABYAI_SPECIFIC_IMPROVEMENTS.md), there's evidence that:

  • Late layers (24-35) may be more important for BABYAI-specific improvements
  • BABYAI tasks (navigation/instruction-following) may benefit more from higher-level reasoning in later transformer layers

Scoring Example

If a model performs well on BABYAI:

Scenario A: Model excels on BABYAI + 2 other environments (wins 5 Layer 3 subsets)

  • Reward: 5 × 0.143 = 0.715

Scenario B: Model excels on BABYAI + 6 other environments (wins 1 Layer 7 subset)

  • Reward: 1 × 16.0 = 16.0

Scenario C: Model excels on ALL 8 environments including BABYAI (wins Layer 8)

  • Reward: 1 × 256.0 = 256.0

Result: Comprehensive performance (Scenario C) gives 358× more reward than partial performance (Scenario A)!

Conclusion

BABYAI is a critical component of the validation system, appearing in:

  • 120 out of 219 evaluated subsets (54.8%)
  • All validation layers 3-8
  • Maximum reward potential of 447.009 (if model wins all BABYAI-related subsets)

To maximize BABYAI-related rewards, models must demonstrate comprehensive ability across multiple environments, with the highest reward (256.0) coming from performing well on all 8 environments simultaneously.