TensionLM-TS-Trace-Distilled-v11

This is the first trace-distillation artifact from the TS reasoner line.

Teacher: TensionLM-117M-TS-Reasoner-v10

Student: a compact CPU-trained TensionLM checkpoint with about 1.1M parameters, trained on compact v10 traces.

What Is Included

• student/latest.pt: compact TensionLM trace student checkpoint.
• student/tokenizer.json: tokenizer trained on trace text.
• data/train.jsonl, data/val.jsonl, data/test.jsonl: v10 trace-distillation rows.
• data/*.txt: LM text format used for student training.
• eval/trace_distilled_v11_training_summary.json: training/eval summary.
• eval/trace_distilled_v11_long_eval.json: raw held-out imitation evaluation.
• generation/eval scripts used for the receipt.

Dataset Shape

Each row contains:

• prompt,
• category,
• distribution,
• v10 teacher trace,
• compact target:

answer: ...
rule: ...
confidence: ...
rationale: ...

The dataset covers six families:

• graph/transitivity,
• arithmetic,
• code traces,
• boolean logic,
• set reasoning,
• string reasoning.

And four distributions:

• standard,
• paraphrase,
• unknown,
• mixed.

Receipts

Dataset:

• total rows: 1920
• train: 1632
• val: 144
• test: 144

Student:

• parameters: about 1.1M
• training steps: 580
• trace-text validation perplexity from training: 2.45

Held-out raw imitation:

• rows scored: 144
• generation rows: 48
• raw exact answer hits: 0/48
• raw exact rule hits: 0/48

Claim Boundary

This is an honest first neural bridge. The compact student learned the trace-text distribution, but raw exact answer/rule generation is not solved yet. The v10 symbolic reasoner remains the working reasoning system. v11 provides the dataset, student checkpoint, and evaluation harness needed to improve neural imitation in the next wave.

Usage

python inference.py --prompt "Logic board: A=true; B=false; C=true. Evaluate A AND NOT B:" --category boolean_logic
python eval_trace_distilled_v11.py --checkpoint student/latest.pt --test_jsonl data/test.jsonl