Skip to content

NAZ LLM EVALS + row_log.py (llama as a judge script) + recent_utils_py (most up to date) - #2

Open
nazcol wants to merge 8 commits into
mainfrom
naz
Open

NAZ LLM EVALS + row_log.py (llama as a judge script) + recent_utils_py (most up to date)#2
nazcol wants to merge 8 commits into
mainfrom
naz

Conversation

@nazcol

@nazcol nazcol commented Apr 4, 2025

Copy link
Copy Markdown
Collaborator

No description provided.

@nazcol

nazcol commented Apr 4, 2025

Copy link
Copy Markdown
Collaborator Author

serve request that is erroring below:

vllm serve meta-llama/Meta-Llama-3-8B-Instruct
--gpu-memory-utilization 0.95
--max-model-len 2048
--port 8001
--tensor-parallel-size 1
--max-num-seqs 32

@nazcol nazcol changed the title row_log.py (llama as a judge script) + recent_utils_py (most up to date) NAZ row_log.py (llama as a judge script) + recent_utils_py (most up to date) Apr 4, 2025
@nazcol nazcol changed the title NAZ row_log.py (llama as a judge script) + recent_utils_py (most up to date) NAZ row_log.py (llama as a judge script) + recent_utils_py + LLM EVALS (most up to date) May 1, 2025
@nazcol nazcol changed the title NAZ row_log.py (llama as a judge script) + recent_utils_py + LLM EVALS (most up to date) NAZ LLM EVALS + row_log.py (llama as a judge script) + recent_utils_py (most up to date) May 1, 2025
lisadunlap
lisadunlap reviewed May 1, 2025
View reviewed changes

@lisadunlap lisadunlap left a comment

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

looks good! Mostly minor things and making sure we are updating files instead of making new ones

Comment thread vibe_baseline.py

@lisadunlap lisadunlap May 1, 2025

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

how does this differ from the get_differences.py file? If we are modifying that file let's name it the same thing so we don't muddy the repo

@nazcol nazcol May 1, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

they actually are doing the same thing just checked now, I can change the name of mine nevertheless though

Comment thread minimodel_responsesnazx1.csv

@lisadunlap lisadunlap May 1, 2025

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we move these to a folder

Comment thread recent_utils_llm.py

@lisadunlap lisadunlap May 1, 2025

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we instead update the existing utils_llm file

Comment thread .DS_Store

@lisadunlap lisadunlap May 1, 2025

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

remove

nazcol and others added 3 commits May 1, 2025 19:57
ethantsliu added a commit that referenced this pull request Jul 20, 2026
@ethantsliu
docs: consolidate results layer + back the overcount (3.5x, corrected…
bd91136 
… mechanism) + self-control + variance

- Overcount CONFIRMED on the current matrix (n=4712, compute_overcount.py): Llama-Guard flags 25.4%
  unsafe vs 7.2% genuine harm = 3.5x overall (6.7x AdvBench). MECHANISM CORRECTED: 79% of Guard's flags
  are false-positive CLEAN REFUSALS (Guard reacts to prompt topic, not the response); it also misses 25%
  of genuine harm. NOT driven by refuse-then-leak (which is a separate real phenomenon, 144 cases both
  judges agree on). RESULTS.md #1 + RELATED_WORK updated to the measured mechanism.
- Imitation-specificity CONFIRMED at seed-42 scale (self-SFT control): ministral imitates ITSELF -0.4pt
  (null, 4/4 ds) vs OTHERS +3.4pt; aya null both ways. Supersedes the pilot-only evidence. RESULTS.md #2.
- variance_decomp.py: reproducible source 79% / target 3% / dataset 0.2% decomposition + stats artifact.
- Docs reconciled: README + RESULTS.md source-conditioned (was contradictory "target-conditioned");
  dissociation table CONFIRMED (9 CLEAN/4 PC_FAILS/1 PC_INVALID/2 deferred); roster-asymmetry rationale
  (steering N=20 vs imitation N=9, matched core=8); new docs/RESULTS_INDEX.md links the 3 result stores;
  data/results/safety/README.md rewritten (was a false "scaffold only" placeholder).
ethantsliu added a commit that referenced this pull request Jul 20, 2026
@ethantsliu
docs(README): state evidence strength per finding; reframe thesis to …
ddbb69f 
…match the data

The README led with 'imitation causes erosion of safety', but the matrix shows
mean erosion of +0.22pt with 245/540 adapters getting safer and only 2 of 9
sources eroding above 1pt. Leading with an effect that small invites the obvious
reviewer question.

Reframes the thesis to what the evidence supports: imitation transfers identity
without transferring safety, and the reported alarm is largely measurement
artifact. Under that framing the guard overcount explains the field's alarm, the
small source-conditioned erosion is the empirical result, and the steering
dissociation is the mechanism.

Adds an explicit evidence-strength table (strong / solid-null / thin) and
downgrades Finding #2 from CONFIRMED to SUPPORTING (n=1): the self-SFT-vs-cross
contrast is meaningful on ministral-8b alone. Also flags that the connecting
claim rests on n=1 within the matched core, and that granite-4-h-small's
outstanding steering verdict is what would take it to n=2.

(cherry picked from commit b1f202dda85627b60a79936f36ac6e19089bf2db)
ethantsliu added a commit that referenced this pull request Jul 20, 2026
@ethantsliu
docs: demote imitation-specificity to appendix; caveat the base-condi…
1bffc68 
…tioning; fix backwards docstring

Realigns the four findings so the framing matches the evidence strength:

- Two PILLARS kept prominent: the measurement overcount (#1) and the causal
  identity/safety separability (#4). Finding 4 restated as a mechanism result
  ('where safety lives') rather than 'weight-imitation erodes safety', which the
  near-null erosion does not support.
- Erosion finding (#2) restated as a SOLID null: mean +0.2pt, 45% get safer,
  imitation does not meaningfully erode safety.
- Base-conditioning (#3) softened to TENTATIVE: the 79/3/0.2 decomposition
  partitions a near-noise signal (harm rates quantised to ~0.5pt, no significance
  test in the single-seed table); 3-seed pilot finds the same pattern with
  p-values. 'Asymmetric laundering' now framed as a conditional security reading,
  not a confirmed result.
- Imitation-specificity (self vs cross, n=1) demoted to an explicit single-source
  supporting observation (appendix-bound), removed from the headline findings.

Also fixes a real bug: variance_decomp.py's docstring had source/target defined
backwards ('source = the model being imitated') and concluded erosion 'tracks WHO
you imitate' -- the opposite of the data. Verified against the CSV: source = the
base model that is fine-tuned; erosion tracks the base, not the imitated target.
Numbers unchanged; only the interpretation was wrong.

Fixes cross-doc inconsistency (Finding 2 was 'CONFIRMED' in RESULTS.md vs 'thin'
elsewhere) and a broken '#5' cross-reference.

(cherry picked from commit 0033ad28bb15cefd727b1f7d74a27697ad79a62c)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@lisadunlap lisadunlap lisadunlap left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@nazcol @lisadunlap