Skip to content

FILE_MANIFEST.md

Latest commit

 

History

History
167 lines (134 loc) · 7.66 KB

FILE_MANIFEST.md

File metadata and controls

167 lines (134 loc) · 7.66 KB
Raw

ABI Open-Source File Manifest

This document defines exactly which files belong in the published ABI repository. Everything not listed here is either LayerCake-specific infrastructure or development-only archive material.

Repo Identity

Repository name: abi-transfer
Tagline: Autonomous Basis Injection — domain behavioral reconstruction across LLM architectures

INCLUDE — Production Package (abi/)

File Purpose
abi/__init__.py Package entry point
abi/models.py AnchorABI, CandidateABI, DomainModule
abi/training.py Stage A (domain), Stage C (KD calibration), Stage D (Procrustes)
abi/evaluation.py NIB evaluation logic

INCLUDE — Entry Points

File Purpose
run_abi.py Full Path 2C training + NIB evaluation (same-backbone T5-large)
verify_result.py Standalone result verifier (no GPU, < 5 s)
wikitext_cache.py WikiText-2 data loader (used by all experiment scripts)
reproduce_abi.py Cross-size paste reproduction harness

INCLUDE — Locked Experiment Scripts (one per validated claim)

File Claim Result file
cross_arch_t5_nib_v53.py Path 2C — T5-large same-backbone NIB PASS cross_arch_t5_nib_v53_results.json
cross_family_nib.py Exp 32 — GPT-2-small → Qwen2.5-0.5B cross-family NIB PASS cross_family_nib_results.json
cross_arch_enc_dec_nib.py Exp 39 — T5-large (enc-dec) → GPT-2-medium (dec-only) NIB PASS cross_arch_enc_dec_nib_results.json
cross_arch_t5_succession.py Exp 40 — T5-large backbone-update invariance (304% efficacy) cross_arch_t5_succession_results.json
scale_validation_test.py GPT-2-medium 354M backbone-update invariance (65% efficacy) scale_validation_results.json
succession_test_v2.py 3-round GPT-2 succession stress test succession_results_v2.json
calibration_scaling_law_b.py Exp 35b — calibration step scaling law (R² ≈ 0.97) calibration_scaling_law_b_results.json
cross_lineage_transfer_test.py Cross-lineage Pythia → GPT-2 transfer cross_lineage_results.json
cross_size_large_nib_v9.py Cross-size 117M → 774M decoder NIB PASS cross_size_large_nib_v9_results.json
multi_domain_atlas.py Multi-domain atlas (Python + WikiText + SQL) multi_domain_atlas_results.json

INCLUDE — Locked Result Files

File Experiment Key metric
cross_arch_t5_nib_v53_results.json Path 2C (45AS) top-5 = 0.8725 PASS
cross_family_nib_results.json Exp 32 top-5 = 0.8701 PASS
cross_arch_enc_dec_nib_results.json Exp 39 top-5 = 0.8699 PASS
cross_arch_t5_succession_results.json Exp 40 efficacy = 304.3% PASS
scale_validation_results.json Scale (GPT-2-medium) efficacy = 65.3% PASS
calibration_scaling_law_b_results.json Exp 35b R² ≈ 0.97
cross_lineage_results.json Cross-lineage PASS
cross_size_large_nib_v9_results.json Cross-size PASS
procrustes_nib_results.json Procrustes NIB baseline reference
analytical_calibration_results.json Analytical calibration reference

INCLUDE — Documentation

File Purpose
README.md Primary landing page — core claim, results table, reproduction
PROOF.md Formal proof structure — protocol, evidence, what remains open
ABI_EXPERIMENTS.md Complete experimental ledger — all 8 validated claims
ABI_ARCHITECTURE.md Architecture spec — tap configuration, d_abi rationale, null-space geometry
ABI_REPRODUCE.md Step-by-step reproduction guide
ABI_START_HERE.md 10-minute onboarding for new contributors
OPEN_SOURCE_README.md Developer reference — component table, honest claims
FILE_MANIFEST.md This file
.gitignore Excludes checkpoints, logs, dev scripts
requirements.txt Python dependencies

INCLUDE — Supporting Experiments (context only)

File Purpose
abi_ablation_test.py Objective ablation (corrMSE vs KL vs logit-MSE)
abi_ablation_results.json Results
knowledge_non_interference.py Domain isolation test
knowledge_non_interference_results.json Results
non_inferiority_benchmark.py NIB protocol reference implementation
non_inferiority_results.json Baseline NIB result
nib_geometry_diagnostic.py ABI geometry analysis
nib_geometry_diagnostic_results.json Results
generation_equivalence_test.py Generation-level equivalence test
generation_equivalence_results.json Results
precision_parity.py fp32 vs bf16 parity
precision_parity_results.json Results
method_robustness_sweep.py Robustness across seeds/hyperparams
method_robustness_results.json Results
ranking_quality_analysis.py Top-k ranking analysis
ranking_quality_results.json Results
transition_zone_multiseed.py Calibration step transition zone
transition_zone_results.json Results
abi_scaling_law.py ABI capacity vs model size
abi_scaling_results.json Results
analytical_calibration.py Analytical step count predictor
calibration_budget_floor.py Minimum calibration floor
calibration_budget_floor_results.json Results

DO NOT INCLUDE — Archive (development only)

Iterative experiment scripts (superseded by locked final versions)

  • cross_arch_t5_nib_v1.py through cross_arch_t5_nib_v60.py (except v53)
  • All corresponding _log.txt and _results.json files
  • cross_arch_path3a_*.py, cross_arch_path3a_*.json, cross_arch_path3a_*.pt
  • cross_arch_nib_v2.py, cross_arch_nib_v3.py
  • cross_size_large_nib_v1.py through cross_size_large_nib_v8.py
  • cross_size_nib.py, cross_size_nib_v2.py, cross_size_nib_v3.py
  • cross_arch_qwen2_v*.py
  • cross_model_transfer.py, cross_model_transfer_v2.py

All run logs and error logs

  • *_run.log, *_err.log, *_log.txt, atlas_run_*.log, atlas_err_*.log

Intermediate model checkpoints

  • All .pt, .ckpt, .safetensors files (except baseline_5k_ckpt.pt if needed for reference)
  • cross_arch_t5_nib_v*_calibrated.pt

Development / diagnostic / patch scripts

  • _fix_gen_test.py, _fix_gen_test2.py
  • _patch_*.py (all patching scripts)
  • check_*.py, debug_*.py, diagnose_*.py, diag_*.py
  • inspect_*.py, compare_*.py (most comparison scripts)

LayerCake LM system (separate repo)

  • layercake_model*.py, layercake_data.py, layercake_decoder.py
  • train_layercake_*.py, train_domains_*.py
  • layercake_domains/
  • cortana*.py, cortical_swarm/, cortex_governor.py
  • autonomous_moa.py, moa_*.py
  • mcp_*.py, auto_router*.py
  • All LAYERCAKE_.md, CORTANA_.md, MOA_*.md docs

Investor / business / pre-publication material

  • INVESTOR_*.md, BUSINESS_MODEL.md, PATENT_*.md, FULL_PLATFORM_*.md

Validated Claim Ladder (summary)

Claim Experiment Top-5 / Efficacy Status
Same-backbone NIB (T5-large) Path 2C / 45AS 0.8725 LOCKED
Cross-family decoder-only (GPT-2 → Qwen2.5) Exp 32 0.8701 LOCKED
Cross-architecture enc-dec → dec-only Exp 39 0.8699 LOCKED
Backbone-update invariance (dec-only, GPT-2-medium) scale_validation_test 65.3% PASS
Backbone-update invariance (enc-dec, T5-large) Exp 40 304.3% LOCKED
Cross-lineage (Pythia → GPT-2) cross_lineage PASS PASS
Cross-size (117M → 774M decoder) cross_size_large_nib_v9 PASS PASS
Calibration scaling law Exp 35b R² ≈ 0.97 PASS