ntc-ai/evidence-dial-behavior

Evidence Dial data and complete experiment record

The latest study is under expanded138-h1024-v2/. It retains 138 accepted original contexts across 37 training questions, two fresh 1,024-update cohorts, all 24 complete-answer DEV settings and five prelocked conditions on 32 fresh confirmation questions. The combined confirmation table has 160 rows, representing 32 matched questions; these are not 160 independent questions. All previous experiment folders are preserved.

Condition (32 fresh questions)	Mean behavior score
OFF	0.686719
Fixed prompt-only	0.649375
PG strength 0.5 — descriptive	0.682734
PG strength 1 — selected on DEV	0.673906
PG strength 2 — descriptive	0.639062

The target is sympathetic, constructive support with independent judgment. The learned adapter receives the same ordinary user question, with no added behavior instruction. A separate fixed prompt-only condition measures what prompting already achieves. Scores are NanoGPT Jev model judgments, not accuracy or human preference rates.

Both fresh cohorts, using 138 contexts across 37 questions, finished 1,024 updates with raw weights and no EMA. Development selection across 24 settings chose G=0.00025, step 256, strength 1 (DEV 0.6950). Longer training did not improve the maximum. All 160 fresh-confirmation responses end naturally; the 4,096-token safety cap was not reached (longest: 575 tokens).

The selected adapter's fresh score is 0.673906 versus OFF 0.686719 and prompt-only 0.649375. ON minus OFF is −0.012812, with an individual 95% paired question-bootstrap interval [−0.047266, +0.020781]; the change from prompt-only is +0.024531 [−0.000547, +0.049687]. Neither establishes an advantage. The selected sympathy diagnostic decreases and sycophancy increases. Strength 2 regresses. Strength 0.5 is descriptive, not a new test-selected winner.

One synthetic corpus, one training seed, shared curation/evaluation judge and limited first-state supervision constrain generalization. The fresh questions were frozen before reading trained expanded-run development results; broader semantic families overlap. Length changes may affect the judgments; no length-matched benefit is established.

The new folder includes the actual 138 paired-state cache and its index, quarter FULL/RAW/NATIVE snapshots at 0/256/512/768/1,024 for both cohorts, exact update journals, scoped initialization/native checks, all predictions, fresh-question manifest and final independent admissions. provenance/observations-and-sources.tar.gz preserves 3,579 archived files, including execution, native-generation, judgment and health records and pinned sources, including original missing-wrapper and known HTTP 401 failures. These operational errors were recovered without replaying completed updates or successful/unknown model judgments. Original host paths are execution provenance; port them through an explicit new configuration.

The fresh 32 questions were never training inputs or checkpoint selectors. The original 32-question behavior test was already spent and is not presented as the fresh confirmation. Curation itself retains its original 128-token sampled replies and ≥3/4 threshold; the new evaluation uses completed answers. Uniform sampling over 138 context pairs gives unequal exposure across 37 questions.

• Thorough README and code
• Raw behavior adapter
• All states, predictions, logs and audits
• Model collection
• Dataset collection

Archived earlier studies

The material below describes the original 37-pair and cutoff-correction records, which remain unchanged.

Evidence Dial behavior data and experiment record

Track 1 section 7 preference/style alignment on Tiny Aya Global 3.35B. Questions are synthetic:64 TRAIN,16 DEV,32 held-out. The original behavior models train on 37 selected context pairs. Each pair keeps the same question in the neutral and positive contexts, adds a behavioral system instruction only to the positive context, and has empty continuation history. Targets are frozen final hidden states before the first response token.

curation/attempts-1024.jsonl preserves all 64 questions×16 instruction trials and model-judged effects, including failures. selected-37.jsonl contains the original max-per-question contexts meeting the frozen 3/4 threshold. passing-138.jsonl retains all 138 contexts meeting that same threshold across the same 37 questions; it is a separate prepared data derivative, not 138 distinct questions and not training data used by the original models. Sampling those context pairs uniformly gives unequal question exposure.

training-pairs contains actual saved positive/neutral states. curation/native-states retains positive first states and sampling RNG state for all 1024 trials. training preserves raw/full snapshots, native exports, exact update journals and scoped initialization records. evaluation contains every finished development and locked held-out arm. audits binds the source, model, data, first observations and decisions. Original host paths in archived provenance identify recorded executions; port them using an explicit new execution configuration.

The complete held-out comparison includes OFF, a fixed prompt-only instruction, all development-selected maximum ties, and a three-strength grid from the display checkpoint frozen before test. The display grid is descriptive; test scores do not select another model. The primary selector is equal-question mean intended-behavior score ON minus matched OFF. Sympathy and sycophancy judgments are separate diagnostics.

Jev is typesafe/jev-1.14 through NanoGPT. Scores are model judgments, not accuracy percentages or human preference rates. Curation uses 128-token sampled replies; all evaluation arms use the same 64-token greedy cap and native EOS, retaining truncation. Weight EMA is disabled; all adapter, critic and bank weights are raw. The small synthetic corpus, one training seed, shared curation/evaluation judge and first-state supervision limit generalization.

secondary-QA contains a separate completed confirmation: PG and SFT tie at 174/192 on DEV, while all five preselected PG arms score 221–229/256, all four SFT controls 230–234/256, and OFF 142/256 on sealed test. This establishes a development tie and held-out gap, not SFT parity. Earlier grounding/FaithEval results remain separate in the public project.

Code: https://github.com/HyperGAN/evidence-dial Adapter: https://huggingface.co/ntc-ai/evidence-dial-behavior-particle Model-derived outputs and states retain CC-BY-NC-4.0; retain upstream code licenses. No credentials or private Git/Babel history are included.

Completed behavior result

The development-selected strength-1 adapter scores 0.599609 on 32 held-out questions versus 0.588828 OFF and 0.612266 prompt-only. Its paired change from OFF is +0.010781, with an individual 95% question-bootstrap interval [−0.012188, +0.034687]. Improvement over OFF is uncertain, and there is no demonstrated advantage over prompting. Strength 2 regresses to 0.520234. The separate sympathy diagnostic decreases at the selected strength. The project does not claim that this study establishes greater caring or that ParticleGAN beats SFT.

The source README contains all 24 development cells, the locked QA comparison, limitations, and complete recorded examples. The dataset's provenance/observations-and-sources.tar.gz preserves 8,555 native/judge/health observations and execution records, including 45 pinned source files; its internal manifest records content hashes.

Complete-answer correction: fixed old checkpoint, DEV only

The original 64-token evaluation cut off useful answers. A separate complete-answer remeasurement of the published 32-update, 37-pair checkpoint covers all 16 development questions, five conditions, and 80 responses. Every answer reaches native EOS; all original saved token prefixes match exactly. The 4,096-token safety cap was never reached (longest response: 601 tokens).

Condition	Mean behavior score	Mean generated tokens
OFF	0.662031	401.75
Prompt-only	0.666563	60.44
Strength 0.5, descriptive	0.684219	256.44
Strength 1, previously selected under the 64-token protocol	0.677500	111.19
Strength 2, descriptive	0.574219	48.44

The previously selected strength 1 gains +0.015469 over OFF, with an individual paired 95% question-bootstrap interval [−0.041875, +0.075004], and +0.010937 over prompt-only [−0.027031, +0.050312]. These are uncertain DEV comparisons. They do not establish an untouched-test advantage. No new winner is selected from this remeasurement; strength 2 regresses. The older 32-question test remains scoped to its 64-token protocol.

Fresh training on all 138 passing context pairs, from 0 through 1,024 updates, is underway on both GPUs. The published adapter files still identify the original 37-pair, 32-update checkpoint; new expanded-run weights and results are pending. New full-answer evaluation requires EOS; the original 128-token sampled mining records are retained.

Complete-answer demo, correction data and raw observations, method and limitations. The recorded full-answer path is evidence_dial.behavior_fullanswer_generate_v2.py; the archived short-answer evaluator is separately identified.