dwidlee / systemone-lite-0.5b

systemone-lite-0.5b

Fine-tuned weights for systemone-lite : typed decisions via next-token scoring over option aliases ( choice / encoded noul / score ).

Not affiliated with TypeSafe AI or Jev.

Chess-specialized weights are a separate local checkpoint; this model is not trained on chess.

Accuracy (option top-1)

Hard: alternate state layouts, option subsets, paraphrases (same label rules). Repo reports: benchmarks/general_*_eval*.json .

Selected iid (this model): ticket.route 1.000, alloc.fund_next 0.980, ticket.needs_human 0.775. Near base: debate.winner 0.493, debate.enough_evidence 0.460.

Chess transfer (same eval harness)

Move top-1 on chess_eval_5k (n=500): base 0.790 , this checkpoint 0.458 , chess-only SFT 0.834 . General SFT does not improve chess.

Latency (inference path; base 0.5B measured)

In-process on RTX 3060 , warmup excluded. Same scoring path this checkpoint uses (0.5B forward dominates latency).

Option scoring vs AR JSON ( model.generate greedy multi-field JSON, full vocab). Source: repo benchmarks/latency_vs_ar.json .

Option path: batched next-token logits; softmax over option token ids; prefix KV. AR runs often used the full max_new_tokens budget (no early EOS). Option path is schema-constrained; AR JSON validity is not guaranteed in this bench.

TypeSafe public materials cite Jev E2E roughly 70–500 ms (cloud + network; not measured here).

Inference method (server)

1. Encode shared state once (prefix KV).
2. Batch per-question suffixes.
3. Softmax only over criteria / yes–no / score-level token ids.
4. Assemble System One–shaped answers .

systemone-lite --model dwidlee/systemone-lite-0.5b --port 8000

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "dwidlee/systemone-lite-0.5b"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

Limits

• Synthetic rule labels; not human preference data or RLCD.
• Option softmax ≠ population calibration (no ECE curves published).
• Multi-token option strings are not first-class (training uses letter aliases).