agentic-moral-alignment/persona-and-other-evals

Qwen3.5-9B AMA adapters — persona evals

Inference code, the data it produced, and the tools that turn that data into tables and an HTML viewer. The evals are Anthropic's persona set, scored in three regimes: teacher-forced logprob of the answer literal, greedy answer with the reasoning block pre-closed, and a full 16k-budget reasoning trace.

Pinned models

• base unsloth/Qwen3.5-9B @ 005429cee5cb648998cf2b70eebdd83175989c9a
• util agentic-moral-alignment/qwen35-9b__ipd_str_tft__util__native_tool__r1__core @ 13f129e0214a26cfcd0b35c0daa4fbea19465692
• deont agentic-moral-alignment/qwen35-9b__ipd_str_tft__deont__native_tool__r1__core @ d2bd11937f2c34994b24f469f78d0fa658352a27
• game agentic-moral-alignment/qwen35-9b__ipd_str_tft__game__native_tool__r1__core @ 5a40fa4c3e88e542a941bf337e649bbaf200ece0

Every run verifies these revisions and the adapter SHA256 before loading; base is the same weights with the adapter switched off via disable_adapter(), measured as exactly equal to the unwrapped base.

Reproduce the inference

export HF_HOME=/path/to/hf_cache
# step 1: teacher-forced logprob + direct answer, one model per GPU, ~50 min
./run_1_fast.sh
# step 2: full reasoning traces via SGLang, 16k budget + forced close, ~15 h
./run_2_thinking.sh
# progress of any run, at any time
./run_3_status.sh
# pooled + per-file rates for all three regimes
python analyze.py --per_file

run_1_fast.sh is resumable per (file, question, sample). run_2_thinking.sh reshards and rewrites its output dirs, so move them aside to keep a previous run. Both verify the pinned revisions and adapter hashes before loading.

Build the HTML visualisation

# reasoning-trace viewer: one page + a per-eval data folder
python viz_traces.py --per_group 15      # results/viz/traces.html
# persona-polarity bar plots
python plot_polarity.py                  # results/plots/*.png

viz_traces.py writes traces.html plus data/meta.js and data/<eval>.js. Open traces.html directly — the data loads through <script> tags rather than fetch(), which is what makes it work over file://; a fetch-based page silently shows nothing there. Keep the page and its data/ folder together. --per_group 0 embeds every question (hundreds of MB); the default samples 15 flipped and 15 unflipped per eval.

Code

eval / inference: run the persona evals on the pinned models

• eval_qwen35_ama_persona.py
• persona_eval.py
• run_1_fast.sh
• run_2_thinking.sh
• run_3_status.sh
• selected_persona_evals.txt
• sglang_thinking_eval.py
• smoke_adapter_effect.py
• track.py

analysis and visualisation of those outputs

• analyze.py
• package_release.py
• plot_polarity.py
• viz_traces.py

Data included

data/ is the eval output itself, about 15 GPU-hours to regenerate:

path	what it is
data/persona_fast_<model>/<model>.logprob.jsonl	per question: logprob of each answer literal, sum / first-token / length-normalised
data/persona_fast_<model>/<model>.gen.jsonl	greedy answer with the reasoning block pre-closed, plus the generated text
data/persona_thinking_shard*/generations.jsonl	one row per (model, question): full trace, answer, parsed verdict, token count, whether the trace was force-closed
data/persona_fast_<model>/rates_*.csv	per-file match rates

Result tables

file	what it is
results/analysis/answers_long.csv	every (model, eval, question, regime) verdict in long form
results/analysis/thinking_flips_per_eval.csv	per eval and adapter: match rates, flips, direction of flips
results/analysis/flip_pct.csv	flip percentage per eval and adapter
results/analysis/coherence*.csv	agreement between the three scoring regimes
results/analysis/timings.csv	wall time per stage

Results of the eval campaign

logprob   base 56.68   util +0.03   deont -0.09   game -0.02   flips 0.6-0.8%
direct    base 57.50   util +0.01   deont -0.04   game +0.17   flips 0.9-1.0%
thinking  base 46.68   util +0.44   deont +0.08   game +0.05   flips 4.8-5.6%
          forced_close  base 7.80%  util 10.78%  deont 8.93%  game 7.39%
          mean tokens   base 2165   util 2706    deont 2350    game 2106

No aggregate persona shift: every pooled delta is under half a point. The adapters do change the reasoning — about 5 % of thinking answers flip, in both directions, and util runs 25 % longer traces than base while looping 10.8 % of the time against base's 7.8 %.

Notes that matter for reuse

• The reasoning template opens <think> inside the prompt, so traces close </think> without ever opening one; treat everything up to the last </think> as reasoning, and render log-prob prompts with thinking off.
• Score the space-less literals (Yes/No): the datasets store them with a leading space, but after a chat template the assistant turn starts flush.
• SGLang runtime LoRA is fast but not repeatable (identical requests reproduced identical text 0-1/6, versus 3/3 under HF/PEFT), so any claim about a specific trace or an exact repeat must come from the HF path.
• The 16k thinking budget is deliberate: cap-hit rate plateaus past it, so the failures are degenerate loops rather than slow reasoning.

Environment the numbers came from

python 3.13.2  torch 2.11.0+cu130  transformers 5.5.0  peft 0.20.0