run directory (provenance + evals)

run/code.patch

@@ -0,0 +1,42 @@
+diff --git a/configs/full_14b.yaml b/configs/full_14b.yaml
+index d04ab7d..31b8bef 100644
+--- a/configs/full_14b.yaml
++++ b/configs/full_14b.yaml
+@@ -2,10 +2,11 @@
+ # 100 open factual questions, mid-depth injection, training strengths to the
+ # measured attack peak. These are the headline runs.
+ #
+-# 48 layers, so 24 is mid-depth. THIS IS A PLACEHOLDER: run the depth and
+-# strength sweeps first (scripts/sweep_layers.py) and set layer/alphas from
+-# what they show, as we did for 3B and 7B. The 7B's behaviour differed from
+-# the 3B's, so mid-depth is an assumption until measured.
++# Layer 24 and the strength range are MEASURED, not assumed. The depth sweep
++# (results/sweep_layers_14b) found layer 24 admits 63/68 viable attacks against
++# 46 at layer 12 and 31 at layer 36, and steers hardest at moderate strength.
++# The strength sweep at layer 24 (results/sweep_alpha_14b) peaks at alpha 1.0
++# (94% steered) and degenerates above 1.2, so training brackets the peak.
+ # Needs a 48GB card: 14B in bf16 is ~28GB before activations.
+ 
+ 
+@@ -31,7 +32,7 @@ steered_frac: 0.7
+ eval_question_frac: 0.2
+ repeats_per_question: 6
+ alpaca_replay_frac: 0.5
+-eval_concepts_per_split: 40    # seeded subsample per split for the eval sweep
++eval_concepts_per_split: {train: 40, heldout: null}
+ 
+ # Training
+ lora_r: 16
+diff --git a/configs/full_3b.yaml b/configs/full_3b.yaml
+index 754685a..e1f1ea2 100644
+--- a/configs/full_3b.yaml
++++ b/configs/full_3b.yaml
+@@ -25,7 +25,7 @@ steered_frac: 0.7
+ eval_question_frac: 0.2
+ repeats_per_question: 6
+ alpaca_replay_frac: 0.5
+-eval_concepts_per_split: 40    # seeded subsample per split for the eval sweep
++eval_concepts_per_split: {train: 40, heldout: null}
+ 
+ # Training
+ lora_r: 16

run/config.yaml

@@ -1,10 +1,7 @@
-# HEADLINE RUN: full 500-concept / 21-category bank on Qwen2.5-3B.
-# Settings fixed by the sweeps (see results/sweep_layers_paper_3b, sweep_alpha_3b):
-#   * layer 18 = mid-depth, the strongest attack surface (worst case for us)
-#   * train alphas extend to 1.2 = the steered-rate peak on the base model
-#   * eval alphas cover past the peak into the degeneracy regime
-# Task = open factual questions; eval concepts subsampled (seeded) to keep the
-# eval tractable at 500 concepts.
+# FULL-BANK run: 500 concepts / 21 categories (5 categories held out entirely),
+# 100 open factual questions, mid-depth injection, training strengths to the
+# measured attack peak. These are the headline runs.
+
 
 model_id: Qwen/Qwen2.5-3B-Instruct
 device: auto

run/run_meta.json

@@ -1,6 +1,6 @@
 {
  "run": "full_3b",
- "started_at": "2026-07-25T12:19:58+0000",
+ "started_at": "2026-08-05T17:16:57+0000",
  "argv": [
   "scripts/run.py",
   "configs/full_3b.yaml",
@@ -81,24 +81,24 @@
   "results_dir": "results/full_3b"
  },
  "git": {
-  "commit": null,
-  "branch": null,
-  "dirty": null,
-  "remote": null
+  "commit": "b16f6e72a9bb0c38dd20d0bab46579325178c793",
+  "branch": "main",
+  "dirty": true,
+  "remote": "https://github.com/JacoDuToit11/steering-resistance.git"
  },
  "env": {
   "python": "3.11.10",
-  "platform": "Linux-6.8.0-49-generic-x86_64-with-glibc2.35",
-  "accelerator": "NVIDIA GeForce RTX 3090",
+  "platform": "Linux-6.5.0-35-generic-x86_64-with-glibc2.35",
+  "accelerator": "NVIDIA A100 80GB PCIe",
   "packages": {
-   "torch": "2.13.0+cu126",
+   "torch": "2.6.0+cu124",
    "transformers": "5.14.1",
-   "peft": "0.19.1",
+   "peft": "0.20.0",
    "accelerate": "1.14.0",
-   "datasets": "5.0.0",
+   "datasets": "5.0.1",
    "numpy": "2.4.6",
    "wandb": null,
-   "huggingface_hub": "1.24.0"
+   "huggingface_hub": "1.26.0"
   }
  },
  "data": {
@@ -111,74 +111,5 @@
    "sha256": "fa2356571420fd8c2a444aee6e8c879b865e0cd21f5ade2938a0308effbed8f2"
   }
  },
- "status": "success",
- "wandb_url": null,
- "headline": "clean 100%->100% · steer_heldout@1.6 correct 0%->3%",
- "finished_at": "2026-07-25T13:02:30+0000",
- "artifacts": {
-  "capability/capability_summary.csv": {
-   "sha256": "d5bbb31095b80e60dd46d2e5cdaa7ef2a781ccc89de5766381b21af1425233b9",
-   "bytes": 128
-  },
-  "capability/capability_summary.md": {
-   "sha256": "ebf7fa98749913e9f77670b2f926ead49fd2dfdf3cacdafbc27b873e088a7218",
-   "bytes": 363
-  },
-  "capability/lmeval_run.log": {
-   "sha256": "f3d93e78d2473755612d80a0afb6eaec8ed8027f13b4545ee8ec4d59370d4731",
-   "bytes": 298468
-  },
-  "capability/m0/gsm8k_cot/Qwen__Qwen2.5-3B-Instruct/results_2026-07-23T18-36-29.229735.json": {
-   "sha256": "fa3feefea52e589b23d972c3393a9222b7affa81d04b782a111cbfb5f6e561bc",
-   "bytes": 14432
-  },
-  "capability/m0/mmlu/Qwen__Qwen2.5-3B-Instruct/results_2026-07-23T18-15-57.670538.json": {
-   "sha256": "dfd686f19edce4cb31bffd091167c599b0596b15987d25303613295099cb4562",
-   "bytes": 139207
-  },
-  "capability/m1/gsm8k_cot/__root__steering-resistance__results__full_3b__m1_resist_adapter/results_2026-07-23T19-07-18.393003.json": {
-   "sha256": "35a97896c6938ea0031d99b63a6787f635f4d88b68a52c7aec53a1849b28a25c",
-   "bytes": 14681
-  },
-  "capability/m1/mmlu/__root__steering-resistance__results__full_3b__m1_resist_adapter/results_2026-07-23T18-41-47.488437.json": {
-   "sha256": "e6e525c0832477c3729aca9663ded0bc63ab583d6c521a550365fb8656f1fc0c",
-   "bytes": 143701
-  },
-  "eval_m0.jsonl": {
-   "sha256": "1ae0cbda4668a35852e1ef234043f320d320ae9cb7cf4017b4d2f79f1427605a",
-   "bytes": 11318115
-  },
-  "eval_m1.jsonl": {
-   "sha256": "1cadc857be9f99e14b2a7c6293f759e7f8f47a0ca1a78e684ab321339f97eb36",
-   "bytes": 9612527
-  },
-  "eval_questions.json": {
-   "sha256": "8771796f4c91a61a901278e799032623900657aaa8f9e44ce789dd1fadd8bf8f",
-   "bytes": 27617
-  },
-  "m1_resist_adapter/adapter_config.json": {
-   "sha256": "390e661694a2f685b6256b5d38a80aa41c5236a115df9cfb5edd4a96da0e568e",
-   "bytes": 1103
-  },
-  "m1_resist_adapter/adapter_model.safetensors": {
-   "sha256": "bc5cf2921ad4f341fc270e0b53641804b4c725c9031d5d232dd6cb1f8877a23d",
-   "bytes": 119801528
-  },
-  "summary.csv": {
-   "sha256": "0d7959f18ba4548ff986d2f9d1d951434a73addb3bafd0822e6153011850a2c3",
-   "bytes": 2621
-  },
-  "summary.md": {
-   "sha256": "cbc9c4aa3d4576d1608186f9e893ae0fc1a94d68655f57ec5126e29b756989c1",
-   "bytes": 2627
-  },
-  "train_examples.json": {
-   "sha256": "3e2ac73debbb5da8cadb4e114a35f6346f015ca0f22d52b9486a7fd2b2b7e824",
-   "bytes": 284577
-  },
-  "vectors.pt": {
-   "sha256": "5c9783914f6419b376d133a3872349ac5a07b42c8456894a3214ffa6618158c4",
-   "bytes": 7757862
-  }
- }
+ "status": "running"
 }