mhough/sot-fig4-countdown-ppo-rollouts

Fig. 4 corpus: every training rollout of a faithful PPO run on Countdown, with LLM-judge verdicts

Part of the Controls & Trajectories programme. This is the complete rollout record of one 250-step PPO run reproducing the configuration behind Fig. 4 of Reasoning Models Generate Societies of Thought (Kim, Lai, Scherrer, Agüera y Arcas & Evans, arXiv:2601.10825), plus every verdict the paper's own LLM-judge instrument produced on a stratified subsample. The paper ships no code and no traces; this is the artifact a replicator needs to check the Fig. 4b/4e claim (that conversational behaviours and persona count rise under accuracy-only RL) against something other than the figure.

What was run

	value
policy	Qwen/Qwen2.5-3B (base, no priming, no SFT)
task	Countdown (TinyZero data build)
algorithm	PPO, verl/TinyZero, full fine-tuning
steps	250 (the paper's horizon), rollout.n=4, train batch 128, mini-batch 64
actor / critic lr	1e-6 / 1e-5, kl 0.001, temperature 1.0, response length 1024
reward	TinyZero Countdown scorer: 1.0 correct, 0.1 format only, 0 otherwise
hardware	2×A100 80GB (RunPod), ~654 s/step
validation reward	0.104 at step 0 → 0.702 at step 250 (peak 0.707)
seed	one; the seed knob is a permutation of the training set (rl/run_seed.py)

Full configuration audit against the paper: docs/paper_fidelity_audit.md and tests/test_fig4_fidelity.py in m9h/societies-of-thought.

Files

rollouts.jsonl — one row per generation printed to the verl log, 2,555 rows: 2,102 on-policy training rollouts (temperature 1.0) and 453 held-out validation generations (greedy, printed at every 10th step). The paper's Fig. 4b is measured on validation generations, so the source field matters: the two series are different distributions.

field	meaning
idx	position in the log
step	the training step whose gradient this rollout contributed to, or the checkpoint a validation generation was sampled from (verl prints both before the step:N line)
source	train (on-policy, temperature 1.0) or validation (greedy, held-out set)
prompt	the Countdown prompt as shown to the policy
response	the policy's generation, cut at its end-of-text token, with the prompt's Assistant: and the scorer's debug echo stripped (see below)
words	whitespace token count of response
sha256	hash of response, the join key for judgments.jsonl

judgments.jsonl — one row per (trace, judge). Every row is a parsed verdict; a judge response that failed to parse was never written as zero.

field	meaning
idx, step, sha256	the rollout
judge_model	the judge that produced the row
prompt_version	paper-v1 = the paper's two prompts verbatim (judge_prompt_behaviours.txt, judge_prompt_persona.txt); v1 = this project's earlier paraphrase, one call
source	as in rollouts.jsonl
question_answering, perspective_shift, conflict_of_perspectives, reconciliation	integer counts, the paper's four behaviours and definitions
n_personas	n_perspectives from the paper's persona prompt (under paper-v1), or the single-call count (under v1)
domain_expertise	the paper's per-perspective expertise profiles, when the prompt produced them

judge_prompt_behaviours.txt, judge_prompt_persona.txt — the paper's prompts (arXiv 2601.10825v1, Supplementary Methods: LLM-as-Judge prompts), with <TRACE> where the trace goes. Note that the persona prompt lists transitional markers ("however," "but," "wait," "actually") as indicators of a perspective boundary.

Known caveats

• Two series. The paper judges held-out validation generations at each checkpoint. The validation rows here are that (greedy, ~17 per checkpoint); the train rows are the on-policy samples PPO trained on. Our first analysis mixed them.
• Scorer-debug stripping. TinyZero's reward function echoes Target: / Extracted equation: / Solution string: A conversation between User and Assistant... before each graded response and its verdict (Invalid equation, No equation found, ...) after it. Both are stripped; the response is cut at <|endoftext|>.
• Judge identity. The paper judges with Gemini-2.5-Pro. The verdicts here are from three Anthropic models. Agreement between them is reported in the repository's results/emergence/cross_judge_*.json, chance-corrected (κ, ICC(3,1)), not exact-match.
• One seed.

Headline, for orientation only

Under the paper's prompts verbatim, with two judges, conflict of perspectives rises from about 1 to 8–10 per trace over training and the persona count from 1.3 to 1.8, on both series: Fig. 4b/4e reproduce. Judge-free, the same traces contain no question marks, the "conflicts" are (not N) rejection lines of an arithmetic enumeration, the second "persona" is always the answer-presentation role, and the pronoun "we" goes from 100% of traces to 0%. An earlier paraphrased prompt (v1, 861 verdicts) found no rise and one persona everywhere; those verdicts are keyed to a previous parse of the responses and are not included here. Read results/emergence/FINDINGS.md §4.10–4.11 before quoting.

Citation

@misc{hough2026sotfig4,
  title  = {Fig. 4 corpus: PPO rollouts on Countdown with LLM-judge verdicts},
  author = {Hough, Morgan},
  year   = {2026},
  note   = {Controls \& Trajectories. https://github.com/m9h/societies-of-thought}
}