license: apache-2.0
language:

• en
  tags:
• moral-foundations-theory
• activation-steering
• mechanistic-interpretability
• llama
• qwen
• mft
• pvq
• ideology
  pretty_name: MFT-LLM Binding-Axis Multi-model Backup
  size_categories:
• n<1K

MFT-LLM Binding-Axis Multi-model Backup (May 2026)

Code, data, and analysis artifacts for ρ-calibrated activation-steering of Moral
Foundations representations in 16 LLM variants, including the binding-axis
extension that tests transfer from MFQ-2 → PVQ-21 + political ideology.

This is a research artifact backup, not a packaged release. Dump of
/workspace/ from the rental box where the experiments ran. Source-of-record
for re-running, re-analyzing, or auditing the runs.

TL;DR

We extract concept vectors (mean-difference of last-token activations) for
five MFT foundations on a model, then steer along them via residual-stream
hooks at one layer. The principal contribution is two-fold:

1. ρ-calibration — α = ρ · r̄_ℓ where r̄_ℓ is the per-layer L2 norm of
   the last-token activation. This makes steering coefficients
   cross-model-comparable.
2. Binding-axis transfer — <foundation>_vs_social_norms vectors fail
   external transfer tests (PVQ values, ideology); the cross-cluster
   care_vs_authority vector (Individualizing − Binding) succeeds. Binding-
   axis transfer test is the headline finding for adapting MFT steering to
   external constructs.

Models covered (16)

Original storage paths (source-of-record)

This repo is a verbatim mirror of /workspace/ from the rental box where
the experiments ran. The repo's relative paths are the same as the absolute
/workspace/<path> on the original machine — every file at
<reporoot>/MFT_LLM_ARR_llama70b/results/... was originally at
/workspace/MFT_LLM_ARR_llama70b/results/....

Tarball note: the high-file-count MFV130/ and projection_results/
directories (~50K small .npy / JSON files total) are stored as
MFV130.tar.gz and projection_results.tar.gz per model directory to
sidestep HF's 128-commits-per-hour limit. After cloning, run
tar -xzf MFV130.tar.gz && tar -xzf projection_results.tar.gz inside each
model directory to restore the original layout.

This is important because:

• Many scripts in MFT_LLM_ARR_<model>/run_*.sh use absolute paths like
  /workspace/model_cache/Meta-Llama-3.1-70B for MODEL_BASE. To rerun, set
  MODEL_BASE to wherever you store the weights.
• Some logs reference /workspace/... paths (e.g. test_report.txt headers
  inside projection_results/). These reflect the original run, not your
  local copy.
• Calibration JSON paths inside steering_calibration.json are the
  rental-box paths; the script reads them via MODEL_BASE env override.

Path conventions on the rental box

Repository layout

.
├── README.md                                    ← this file
│
├── steering_summary_all_models.txt              ← cross-model k_α / k_ρ / L* / r̄_ℓ tables (16 models)
├── projection_full_raw_all_models.txt           ← full per-layer Wasserstein projection reports
├── mmlu_safety_regime.pdf                       ← MMLU accuracy vs ρ on 3 instruct models
├── mmlu_safety_summary.txt                      ← textual MMLU summary
│
├── v4_binding_axis_pkg/                         ← v4 binding-axis pipeline source package
│   ├── README.md                                  full method writeup
│   ├── scripts/                                   PVQ + ideology steering, transfer matrix, binding-axis analysis
│   └── data/                                      MFQ-2, PVQ-21
│
├── MFT_LLM_ARR_llama70b/                        ← Llama-3.1-70B replication (was /workspace/MFT_LLM_ARR_llama70b/)
│   ├── MFV130.tar.gz                            🗜 vectors_npy: 30 dirs × 80 layers × 5 components × 2 (raw + normalized) — extract: `tar -xzf MFV130.tar.gz`
│   ├── projection_results.tar.gz                🗜 split by model variant; 5 vs_social_norms + 10 between-foundation per variant — extract: `tar -xzf projection_results.tar.gz`
│   ├── results/                                 v3 MFQ JSON (10) + v4 PVQ-21+ideology JSON (12) + binding-axis v3 JSON (2) + MMLU safety JSON (1)
│   ├── analysis/                                binding_axis_transfer_analysis.json + transfer_matrix_*.json
│   ├── visualizations/                          k_α / k_ρ slope curves + Δscore vs ρ + ridge + Wasserstein
│   ├── steering_calibration.json                per-layer r̄_ℓ for both variants
│   ├── logs/                                    every phase's stdout/stderr
│   ├── *.py                                     full pipeline source
│   └── run_*.sh                                 phase scripts
│
├── MFT_LLM_ARR_qwen2.5_32b/                     ← Qwen2.5-32B replication (same structure; MFV130.tar.gz + projection_results.tar.gz)
├── MFT_LLM_ARR_qwen3_30b_a3b/                   ← Qwen3-30B-A3B (MoE) replication (same structure; MFV130.tar.gz + projection_results.tar.gz)
│
├── run_three_models_master.sh                   ← orchestrator that ran the new-model batch
├── run_remaining.sh / run_qwen3_remaining.sh    ← resume scripts after partial failures
├── run_mmlu_safety.sh / plot_mmlu_safety.py     ← MMLU safety-regime sweep
└── three_models_master.log                      ← unified log of the 38 h batch run

Pipeline phases (per model variant)

1. Extract     15 foundation pairs × 200 samples × full 80-layer monitoring → vectors_npy/ × 30 dirs (15 × 2 variants)
2. Projection  15 pairs × N MFRC test scenarios → Wasserstein-D / direction tables per layer
3. Calibrate   r̄_ℓ on MFQ-2 prompts → steering_calibration.json
4. Steer (v3)  20 layers × 7 ρ × 5 rollouts × 36 MFQ-2 items, batched bf16 → results/{kind}_<f>_vs_social_norms.json
5. Steer (v4)  same grid + PVQ-21 batched + ideology single → results/{kind}_<f>_externals.json
6. Analyze     binding axis (sign test) + 5×16 transfer matrix

The "v3 batched" steering script (moral_steer_all_layers_batched_v3.py)
collapses the per-text per-layer single forward into a single batched
forward over 36 MFQ-2 items, with adaptive OOM batching, atomic incremental
saves, resume, and a bf16 batched-vs-sequential pre-check. This is what
made the full 16-model sweep feasible on 2× B200.

Headline results

Binding-axis transfer test (care_vs_authority vector, sign-test on 11 outcomes)

Key findings:

• Llama-70B passes both variants, including instruct.
• Qwen2.5-32B-Base is the strongest binding axis (perfect 11/11).
• Qwen-Instruct fails (consistent with Qwen-7B-Instruct from prior work) —
  RLHF appears to selectively neutralize political ideology under steering
  while leaving PVQ values directionally intact.
• MoE Qwen3 fails on both variants — expert routing fragments the
  binding-axis representation; L*=0 fallback signals no eligible deep layer.

MMLU safety regime (ρ ∈ [-0.4, +0.4])

Capability remains within ±1pp of baseline across the entire ρ ∈ [-0.4, +0.4]
range — no degradation at the higher ρ values where prior work suggested
capability would erode. ρ ≤ 0.2 is conservative; the true safe regime is
substantially wider on the binding axis. (Qwen3 baseline ≈ random; chat-template
formatting issue under investigation, but Δ remains small.)

How to use this backup

1. Replicate a steering result — cd MFT_LLM_ARR_<model>/, install env from
   environment.yaml, set MODEL_BASE to a directory holding the model
   weights (download separately from Hugging Face). Then bash run_06_v4_increment.sh rebuilds the v4 results from existing extraction +
   calibration. Or run the full pipeline from run_all_auto.sh.

2. Re-analyze — Load analysis/binding_axis_transfer_analysis.json or
   analysis/transfer_matrix_*.json directly. Schemas documented in
   v4_binding_axis_pkg/README.md.

3. Compare across models — steering_summary_all_models.txt has the
   cross-model k_ρ table (cross-model-comparable, ρ-calibrated). The
   projection_full_raw_all_models.txt has the per-layer Wasserstein
   evidence for vector quality.

What's NOT in this backup

• Model weights (/workspace/model_cache/, ≈ 500 GB) — re-download from
  Hugging Face: meta-llama/Llama-3.1-70B[-Instruct],
  Qwen/Qwen2.5-32B[-Instruct], Qwen/Qwen3-30B-A3B[-Base].
• Conda environment (/workspace/miniforge3/, 1.3 GB) — recreate from
  environment.yaml (Python 3.10, PyTorch cu128 for Blackwell GPUs).

Hardware used

• 2× NVIDIA B200 (180 GB each, 360 GB total VRAM)
• 1.9 TB system RAM, 192 CPUs
• 800 GB local storage (overlay), 24 TB shared
• Total wall-clock for 16-model batch: ~38 hours

Citation

If you use this data, please cite the underlying methods:

• MFQ-2 items: Atari et al. 2023, Moral foundations questionnaire-2, JPSP.
• PVQ-21: Schwartz 2003, A proposal for measuring value orientations across
  nations, ESS Core Questionnaire.
• Ideology single-item: Graham, Haidt & Nosek 2009, JPSP.
• ρ-calibration + binding-axis steering: this work (in preparation).

Contact

Issues / corrections: open an issue on the dataset repo or email the
corresponding author.