Sev-2B-preview

A 2B model that makes multiple-choice decisions in one forward pass, with no reasoning tokens. This is a preview; a full release and a tech report are coming.

• Base: Qwen3.5-2B-Base, all parameters fine-tuned · by Sora · weights under CC BY-NC 4.0
• Output: one probability per lettered option, read from the next-token logits after Answer:
• Fast mode: stop after layer 16 of 24, 1.5x faster, accuracy within about a point

Tree Forcing

A reasoning model (the teacher) writes a chain of thought for each training problem. We turn it into a tree of intermediate judgments and train the student on every node, with the true parents' conclusions written into the prompt: teacher forcing on the reasoning graph instead of on tokens. At inference only the final question is asked.

The released weights are the fine-tuned model blended with the base (alpha = 0.7, chosen on held-out selection sets).

Results

Every model scored on the same items by the same harness; accuracy.

• Ahead of all three comparators (paired bootstrap, 95% interval above 0) on the Knights & Knaves diagnostic, xkk and ProverQA. Behind on JevBench and BBH.
• One training seed; the Knights & Knaves diagnostic comes from our own generator (held-out programs).

Use

# hf download sora42y/Sev-2B-preview sev_infer.py --local-dir .
from sev_infer import decide, load, load_exit_map

model, tok = load("sora42y/Sev-2B-preview")
problem = (
    'Ticket: "I was charged twice for order 5521. Please refund one of the charges."\n\n'
    "Question: Which team should handle this ticket?\n"
    "Options:\n"
    "A) billing: Charges, refunds and invoices.\n"
    "B) shipping: Delivery dates and lost parcels.\n"
    "C) technical: Log-in problems and app errors."
)
decide(model, tok, problem)                                                  # {"A": ..., "B": ..., "C": ...}
decide(model, tok, problem, exit_map=load_exit_map("sora42y/Sev-2B-preview", model))   # fast mode

• Prompt: the problem ending in its Options: block, then \n\nAnswer: (added by decide). No chat template.
• Batching: right-pad and read each row at its last real token. Never left-pad: the linear-attention layers ignore the attention mask.
• Command line: python sev_infer.py --model sora42y/Sev-2B-preview --problem-file problem.txt [--fast]

Fast mode

A100 80GB, batch size 1, eager PyTorch, bf16:

The layer-16 state goes through a tuned lens (tuned_lens_L16.safetensors, fitted without labels) and the LM head. Learn-then-Test certifies layer 16 on held-out items: at most 5% disagreement with the full model (measured 4.1%); over eight evaluation sets the mean accuracy change is +0.1 points.

Calibration (optional)

Divide the letter logits by T before the softmax: T = 1.05 for most questions, 2.00 for ordinal ratings, 0.95 for yes/no decisions about a stated case (plus +0.25 on the Yes logit). decide(..., temperature=T, yes_letter=..., yes_bias=...).

Data and limits

• 84,939 training items from 1,075 sources: public datasets, our own program generators and exam-style multiple choice. Every source and its licence: data_card.csv. Some sources are non-commercial, hence CC BY-NC.
• Teacher: DeepSeek's API model; no teacher output is distributed.
• Evaluation items sharing a 13-gram with the training data were removed before scoring. Related training families (LogiQA, LSAT, intent datasets) are listed in details.md.
• English only. Not for text generation, questions without options, or high-stakes decisions without human review.
• The full card, with every table, the training settings and all disclosures: details.md.

Citation

@misc{yang2026sev2bpreview,
  title  = {Sev-2B-preview: a one-pass decision model trained with Tree Forcing},
  author = {Haotian Yang},
  year   = {2026},
  url    = {https://huggingface.co/sora42y/Sev-2B-preview}
}