axondendriteplus/decipop-v7-r2-multi-entity

decipop · V7 r2 · multi-entity

Multi-entity curriculum treatment. A full-fine-tuned SmolLM2-360M backbone with a linear three-class decision head over the last non-padding hidden state.

Repository · Dataset · Results

Decisions

Given facts, universally quantified rules and a statement, predict TRUE (derivable), FALSE (explicit negation derivable), or UNKNOWN (neither). UNKNOWN is a semantic class, not a confidence-based refusal. No chain-of-thought or proof is generated.

Run on CPU

Use Python 3.12 and the repository's locked environment:

git clone https://github.com/adc77/decipop.git
cd decipop
uv sync --extra train --extra analysis --extra baseline
uv run python examples/predict.py --model axondendriteplus/decipop-v7-r2-multi-entity

The example downloads this custom inference bundle, checks its SHA-256 manifest, and predicts the exact held-out prompts in examples/prompts.json. Use --revision <commit> to pin a release. The standalone bundle also supports python infer.py --input prompts.json after installing requirements.txt. Inputs must match prompt-template.txt; the maximum length is 512 tokens and overlong inputs fail rather than truncate. This is not a stock chat model or a Transformers pipeline checkpoint; use the bundled loader. Model weights are FP32.

Training and selection

• Study: V7, repetition r2, seed 947; dataset config v7-r2.
• Selected update: 900; all arms completed 1,134 updates and 36,288 presentations.
• AdamW: backbone 2e-5, head 1e-4, weight decay .01, clip 1, warmup 5%.
• Effective batch 32, microbatch 8; training and aggregate inference use BF16 autocast.
• Candidates: 0, 250, 600, 900, 1,134. Temperature-fit and selection families are distinct; choose minimum calibrated selection NLL, then fit a new temperature on separate calibration families before opening final tests.
• V7 starts from V4's mixed head; V8 starts from V7 r1's selected treatment. V7's curriculum comparison changes token exposure; V8 matches tokens and pairs.

The pretrained base revision is f8027fd0eaeea54caa13c31d31b9fdc459c38b49. protocol.json, selection-freeze.json, calibration.json, and evaluation.json retain the exact original records. public-release.json links the historical artifact to this public packaging; weights, tokenizer and inference code are unchanged.

Evaluation

V7's treatment-minus-control complete-world gains replicate in both fresh runs. This selected checkpoint is one arm, not a separate independent replication.

Fresh test	Question accuracy	Complete families
test_multi	86.53%	68.06%
test_multi_transformations	82.96%	10.00%
test_multi_depth	86.57%	67.59%
test_multi_wording	75.46%	50.00%
test_multi_signed_chain	86.81%	65.74%
test_original	93.78%	81.67%
test_counterfactual	99.56%	98.67%
test_depth	83.33%	68.00%
test_wording	51.33%	3.00%
test_single_branch	53.50%	11.00%

Complete families require every included question correct. For test_multi this means all people; for test_paired all people before and after the assigned edit; for transformations every included variant and query. These endpoints differ. Intervals in the report describe world sampling conditional on trained models, not a population of training seeds or a joint confidence guarantee.

Limitations

Synthetic, narrow rule distributions; shared templates and vocabulary remain. Missing decipops, distractors and wording changes still cause failures, sometimes confident ones. Global temperature scaling does not guarantee transfer calibration. The weights are research baselines, not validated access-control, legal, financial, medical, freight or general business decision systems. Symbolic names are generated tokens. The dataset contains no real records about those people.

Verification and attribution

Historical private release passed a fresh pinned-download CPU check on eight FP32 reference prompts. This public packaging has its own manifest and release receipt; aggregate published scores are from the original BF16 runs, not new CPU test scores. Use verification.json for reference logits, decisions and numerical tolerances.

Apache 2.0; retain LICENSE and NOTICE. The pretrained model, tokenizer, dependencies and project-authored code retain their attribution. No architecture novelty or performance ranking against other model families is claimed.

Dataset normalization

The linked decipop-worlds release normalizes one entity name to EntityA. This checkpoint was trained and evaluated on the original text. All reported scores refer to those original inputs, not the normalized derivative; name changes can change tokenization and predictions. Labels and proof depths are unchanged. The original dataset is retained privately; see the dataset card for provenance.