qwen3-0.6b-rlcd

A 0.6B option-scoring model. You send a state and one or more typed questions, each with its own options and criteria. For every question it returns a calibrated probability over your options, plus a separate confidence that the answer is among them at all. One forward pass, no generated tokens, runs on a CPU.

• Independent questions: a block-diagonal mask keeps each question from attending to the others, so adding or removing a question never changes the rest.
• Only your options: each option is scored from its own hidden state; the softmax covers your options and nothing else.
• Criteria are read: option keys are arbitrary identifiers; the criteria text carries the meaning.
• Proper-scoring training: cross-entropy against observed outcome frequencies (soft targets where they exist). No RL, despite the family name RLCD.

Quick start

pip install torch transformers

from transformers import AutoModel

model = AutoModel.from_pretrained("thefloydd/qwen3-0.6b-rlcd",
                                  trust_remote_code=True).eval()

response = model.score(
    state={"channel": "email",
           "subject": "Deploy failing since 4.2.1",
           "body": "Since upgrading to 4.2.1 our nightly deploy fails at the migration "
                   "step with 'relation users_pkey already exists'. Rolling back to "
                   "4.2.0 fixes it. We have 200 seats and this blocks our release on "
                   "Friday.",
           "plan": "enterprise"},
    questions={
        "kind": {
            "type": "choice",
            "instructions": "What kind of request is this?",
            "criteria": {"bug":      "Reports something that is broken",
                         "question": "Asks how to do something",
                         "feature":  "Requests new functionality",
                         "billing":  "About payment or invoices"}},
        "needs_engineer": {
            "type": "noul",
            "instructions": "Does resolving this require someone who can read the codebase?",
            "criteria": {"true":  "Requires engineering investigation",
                         "false": "A support agent can resolve it from documentation"}},
    },
)

kind = response["answers"]["kind"]
print({k: round(v, 3) for k, v in kind["probabilities"].items()}, round(kind["confidence"], 3))
# {'bug': 0.465, 'question': 0.497, 'feature': 0.032, 'billing': 0.005} 0.718
print(round(response["answers"]["needs_engineer"]["noul"], 3))
# 0.686

Use the whole distribution: here question and bug are close, so the top choice alone is not a safe answer.

Interface

• The model never sees the question id; put the full question in instructions.
• Up to 255 options per choice question.
• Requests up to 8,192 tokens (state plus all questions). Longer requests are refused, not truncated; split them — questions are independent, so the answers are the same.
• fp32 by default; dtype=torch.bfloat16 halves memory and moves the example slightly (bug 0.455, question 0.508). model.to("cuda") for GPU.
• config.json sets temperature 1.0 and confidence_mode: head; score() applies both.

Evaluation

Internal held-out split: 329 task schemas and 79,396 questions the model never trained on, split by whole task (and by whole structure for generated tasks). 95 % bootstrap CIs.

• Selective prediction: 4.9 % of answers given with p ≥ 0.9 are wrong; that covers 53 % of questions. The confidence head ranks errors better than max probability (AURC 0.064 vs 0.077 on the dev split).
• Rule application: 0.973 on trained rule combinations, 0.945 on held-out ones.
• Abstention: mean confidence 0.96 when the record states the claim, 0.21 when it does not mention it.
• Previous release, same questions: accuracy 0.516, NLL 1.152, 13.3 % wrong at p ≥ 0.9.

Limitations

• Strongest on structured decisions (rules, records, pages, diffs, policies: 0.90–1.00). Plain text classification is 0.54, and NLI, reasoning puzzles, logs, scheduling and multi-hop questions are 0.42–0.59.
• Calibration varies by domain: ECE is 0.12–0.15 on reasoning, security triage, multi-hop and document structure, and 0.33 on log lines.
• It cannot answer off-list. If the right answer is missing, only confidence is meaningful.
• Trained on states up to 2,304 tokens; longer inputs are unmeasured.
• English only.

Training

Qwen/Qwen3-0.6B-Base with LoRA (r 32, alpha 64, dropout 0.05), merged. 8,000 steps, batch 32, lr 2e-4, cosine; checkpoint chosen by out-of-distribution dev NLL. The confidence head is trained afterwards against the frozen scoring model.

Data: ~507k questions over 2,456 tasks in 20 domains — published label schemas, rater-share affect and safety labels, forecasts with published probabilities (GEFS, SPF, Manifold), claim verification (FEVER, VitaminC), NVD, GitHub issues, NLI, legal text, tables, document structure, logs, and code-labelled rules, pages, sensors, tool calls, entity resolution, diffs, scheduling and multi-hop records. The Decision Index benchmark is excluded (its datasets by name, and every state checked against its items). Source licences include Manifold (non-commercial AI training), LogHub (research use) and FEVER / VitaminC (CC BY-SA 3.0).

Citation

@software{qwen3_0_6b_rlcd_2026,
  title  = {qwen3-0.6b-rlcd: a prefill-only option-scoring model for calibrated decisions},
  year   = {2026},
  note   = {Qwen3-0.6B-Base + LoRA, block-diagonal question independence,
            proper-scoring fine-tuning, separate confidence head}
}