Qwen2.5-7B GCAD steering controls
Portfolio summary
Trait-steering experiment: inference-time GCAD banks induce profanity, impoliteness and sarcasm on the pinned Qwen2.5-7B-Instruct checkpoint. In the 32-question sarcasm confirmation, 29/32 steered answers scored at least 50/100 versus 0/32 baseline; coherence fell from 89.59 to 70.72. The separate joint study retained 0/5 combined candidates. Individual effects, quality costs and coupling are reported together. Banks, runtime and measured results are unchanged by this documentation update. The joint manifest is refreshed for the updated card.
This artifact is part of tuneLLM, a public research portfolio covering quantization, LoRA/QLoRA, refusal steering and trait steering. Results and saved examples Β· Reproduction and environment setup Β· Methods and references.
Related releases: GGUF, NF4 base, physics QLoRA, trait controls, refusal experiment. Each release has its own verified environment; activation steering and adapter training use different mechanisms.
Separate 27B work: the text-only NF4 load/save/reload smoke is complete on Qwen3.8-27B, using two simple prompts. It is a functional check, not a quality benchmark. A one-step cache check exceeded its numerical tolerance; that result is retained. The NF4 checkpoint remains local; 27B sarcasm steering and QLoRA training are planned. Recorded scope.
Dynamic inference-time banks for individual profanity, author impoliteness and
sarcasm controls on the pinned
Qwen2.5-7B-Instruct base.
These are activation-intervention banks, not base-model weights or LoRA adapters.
Base weights are obtained separately. This bundle contains banks, runtime and
aggregate evaluation; raw response logs and private grader records are not bundled.
Project code, protocols, compact results and selected examples are available in
the public repository.
Historical source tag steering-controls-v1 and commit
199cf41ac9db9024b14c9afddaea8e060af5b14b identify the original implementation.
| 7B controller | Own evaluation design | Trait baseline β GCAD (random) | Coherence baseline β GCAD (random) |
|---|---|---|---|
| PF, b = 5 | 32 ordinary held-out questions, greedy cap 160 | 0 β 31.125 (0); 14/32 lexicon hits | 85.34375 β 82.6875 (85.6875) |
| SARC, b = 2 | 32 ordinary held-out questions, greedy cap 160 | .0625 β 63.78125 (0); 29/32 β₯ 50 | 89.59375 β 70.71875 (87.71875) |
| Impolite, b = 3.5 | 60 author first replies on 13 unique questions, sampled cap 500 | 1.61667 β 77.36667 (not measured) | 82.68333 β 75.9 (not measured) |
The designs and gates differ. Impolite passed its author qualitative contrasts, with no original random arm; PF/SARC passed their registered held-out gates. SARC has a substantial 18.875-point coherence loss and only 1.125 points of gate margin. Its impoliteness mean 38.75 and correlations are descriptive, not independence or causality evidence. One Claude judge family and reconstructed PF/SARC inputs limit validation; coherence is not correctness or helpfulness. These controls intentionally induce potentially offensive style. Apache-2.0 is the licence; no additional use restriction is imposed here.
Validation
The inference core is byte-identical to validated staging 02: 3 real pinned base zero-coefficient checks, 6 forwards total. No staging 03 GPU rerun is claimed. Current manifest validation supersedes historical deferred fields in archival profiles. METHODS records the repair history and scope.
Pinned local inference
Requirements: Python3.11, torch==2.6.0+cu124, transformers==4.52.3,
accelerate==1.7.0, safetensors and tqdm.
Use base Qwen/Qwen2.5-7B-Instruct revision a09a35458c702b33eeacc393d103063234e8bc28 from a local snapshot whose basename equals
that commit. The example uses local-files-only and no remote model code.
Tokenizer must use EOS151645 as PAD and left padding. Base weights are fetched
separately from their upstream repository, not supplied here.
Choose an immutable HF_REV of this bank repository. Read that revision's
commit message Release manifest SHA256: H and use H as the expected
manifest hash. This verifies file consistency for the pinned revision you choose;
it does not independently prove who published them. Publisher authenticity
rests on the HF account. Download this repository at HF_REV, never an unpinned moving
revision. No GitHub checkout is required to run the downloaded bundle. Use real control files (the loader
rejects control symlinks); a pinned local-directory download provides them:
import os
from huggingface_hub import snapshot_download
PINNED_HF_SHA = os.environ["HF_REV"]
snapshot_download("RenormalizationTech/qwen2.5-7b-gcad-steering-controls",
revision=PINNED_HF_SHA, local_dir="controls")
Then from the downloaded bundle root, with H from that commit message in BANK_MANIFEST_SHA and the pinned upstream base snapshot in LOCAL_BASE_SNAPSHOT:
cd controls
python -m runtime.example --snapshot "$LOCAL_BASE_SNAPSHOT" --trait profanity --manifest-sha256 "$BANK_MANIFEST_SHA" --zero-selftest
python -m runtime.example --snapshot "$LOCAL_BASE_SNAPSHOT" --trait profanity --manifest-sha256 "$BANK_MANIFEST_SHA" --prompt "Describe a rainy afternoon in two sentences."
Use impolite or sarcasm for the other isolated banks. The saved settings are
bank/task/decoding-specific, not comparable doses or jointly validated controls.
The standalone inference interface runs from this bundle. The public GitHub
repository documents the experimental pipeline; some archived raw runs, judge
records and local prerequisites remain outside the public release. See the
reproduction guide
for portable checks versus historical full-study reproduction.
Related outcomes
Methods, source attribution and refusal closures, 4B aggregate, and 4B closeout are included. All 3 fixed-transfer 4B traits were induced, but none met the quality gates. No 4B bank or new 4B HF steering repository is published. This package contains no refusal controller. Outcomes are model-specific; checkpoint differences do not establish a model-size effect. No 4B retuning or joint claim; separate 7B joint outcomes are below; no independence claim. The physics adapter and NF4 base are now public and separate from this package. Original GCAD and Persona Vectors attribution is preserved in LICENSE/NOTICE; public author sources are listed in the included method notes.
Additive joint-controls release
The original release_manifest.json is byte-identical and its full list describes
the original revision 763865908e5e94a020606b31ab66185360f7dd44, not the updated root README/METHODS.
Its eight original critical bank/profile/core/loader hashes remain valid, so
the isolated loader still works at this HEAD. The new joint manifest covers
current cards, joint files and shared payload files. Read BOTH lines from this
pinned HF commit message: Release manifest SHA256: H1 and Joint manifest SHA256: H2.
H1 is the original manifest; H2 is joint_controls/manifest.json. Optional root
.gitattributes and .git metadata, downloader .cache and Python pycache are
not payload files and are excluded only from consumer inventory checks. Root
preserves any Hub-managed .gitattributes and verifies it separately unchanged. These check
file consistency for the pinned revision you choose, not independent publisher identity.
Joint guide, all 12 arm metrics, and sanitized aggregate, and post-hoc lexical audit report all outcomes, paired intervals, calibration/duplicate noise, cross-effects and confounding limits. PF means friendly-rubric swearing, not literal profanity incidence: a lower friendly PF score does not establish suppressed swearwords. The lexical addendum is explicitly post-hoc, using an unchanged narrow proxy; all original guide/gate verdicts stay unchanged. No new base weights, LoRA, raw questions, answers, token arrays or private grading reasons are included. The NF4 base and PRL Writing QLoRA adapter are now public; they are separate artifacts and are not bundled here. This package contains no refusal controller.