qwen3.5-0.8b-open-jev-decision

Typed decision model fine-tuned from unsloth/Qwen3.5-0.8B on ZefanCai/Open-Jev (release-v2-redistributable) with Unsloth FastDecisionModel / DecisionTrainer. For each record (a JSON state plus one or more questions) it returns one probability per allowed option for every question: noul → P(yes), choice → distribution over options, score → distribution over ordered levels.

Architecture

• Backbone: unsloth/Qwen3.5-0.8B (revision 23c69c53358a07516b5827588b3fdb12ae78fd65), loaded in 4-bit NF4 (bitsandbytes), LoRA on the language-model linear layers; vision tower frozen.
• LoRA: r=16, alpha=16, dropout=0, target modules: (?:.*?(?:language|text).*?(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?\.(?:qkv|proj|linear_fc1|linear_fc2|out_proj|in_proj_qkv|in_proj_z|in_proj_b|in_proj_a|gate_proj|up_proj|down_proj|q_proj|k_proj|v_proj|o_proj))|(?:\bmodel\.layers\.[\d]{1,}(?!.*\.experts\.\d)\..*?(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?\.(?:(?:qkv|proj|linear_fc1|linear_fc2|out_proj|in_proj_qkv|in_proj_z|in_proj_b|in_proj_a|gate_proj|up_proj|down_proj|q_proj|k_proj|v_proj|o_proj))).
• Decision head: Clef joint schema head (fp32), config {"hidden_size": 1024, "width": 512, "routing_layers": 2, "layers": 4, "heads": 8, "feedforward": 2048}. Questions of a record are scored jointly; options are read from the prompt (span pooling + lexical embedding prior), so the option set can differ per question.
• Trainable parameters: 38,147,588 (LoRA 10,822,656, head 27,324,932).
• Max sequence length: 2048 tokens per record.

Training data

• Dataset: ZefanCai/Open-Jev, config release-v2-redistributable, revision c67699e13d0ae25e35b77165a4b6b079bedc8aba.
• Model input: state_json (decoded), question, kind, options. target is the supervision only. id, group_id, split, source, metadata_json (incl. teacher/control labels), record_json and original_line_number are never part of the input.
• Rows sharing (group_id, state) are packed into records of ≤ 16 questions and ≤ 2048 tokens; option order is preserved.

Splits are the dataset's own; group_id sets are disjoint across splits. Train is used for training, validation for checkpoint selection, calibration once for temperature fitting, test/OOD for reporting only.

Training procedure

• Loss: soft-target cross-entropy over each question's options (noul = binary log loss on the [no, yes] target; choice/score = cross-entropy against the full target distribution), averaged over questions. label_smoothing=0.0, brier_weight=0.0, ordinal_weight=0.0.
• Optimizer: AdamW, lr 0.0001 (LoRA) / 0.0001 (head), cosine schedule, warmup 50 steps, weight decay 0.01.
• Batch: 2 records × 4 accumulation = 8 records/step; bf16 autocast; Unsloth gradient checkpointing.
• Steps: 2378 (1.000 epochs of 2378 steps/epoch).
• Checkpoint selection: lowest validation eval_loss (0.4202 at checkpoint-2378).
• Seed: 3407.
• Hardware: NVIDIA GeForce RTX 5060 Ti (15.9 GiB); train time 72.3 min; peak CUDA memory {'max_allocated_gib': 2.08, 'max_reserved_gib': 2.85}.
• Software: unsloth 2026.10.3, transformers 5.17.0, torch 2.11.0+cu128, peft 0.18.1, bitsandbytes 0.49.2.

Calibration

Per-kind temperatures fitted once on the calibration split (soft-target cross-entropy, range [0.5, 5]): {"noul": 1.1281152314853171, "choice": 1.0897223768764925, "score": 1.5317417263480422}. They are stored in unsloth_decision_config.json (temperature = [choice, score, noul]) and applied by FastDecisionModel.predict.

Evaluation

Metrics per decision kind (never averaged across kinds). top1_accuracy counts any option tied for the highest target probability as correct; ece uses 15 bins; mae_expected_level is in level units.

validation (3,723 questions)

choice

noul

score

test (10,356 questions)

choice

noul

score

ood (15,701 questions)

choice

noul

score

Usage

from unsloth import FastDecisionModel

model, processor = FastDecisionModel.from_pretrained("thealper2/qwen3.5-0.8b-open-jev-decision", max_seq_length=2048, load_in_4bit=True)
answers = FastDecisionModel.predict(model, processor, state, {
    "q1": {"type": "noul", "instructions": "Will moving up cause a collision?"},
    "q2": {"type": "choice", "instructions": "Choose a direction.", "criteria": {"A": "up", "B": "right", "C": "down"}},
    "q3": {"type": "score", "instructions": "Rate the risk.", "criteria": ["low", "medium", "high"]},
})

Choice criteria use ids A, B, … in the original option order (Clef sorts choice ids); noul options map false/true to no/yes. src/inference.py in the training repository does this conversion.

Files

• adapter_model.safetensors, adapter_config.json: LoRA weights for the backbone.
• joint_head.safetensors, joint_head_config.json: decision head.
• unsloth_decision_config.json: layout, max length, base model, calibrated temperatures.
• tokenizer / processor files of the base model.

Limitations

• Trained and evaluated only on Open-Jev's controlled, mostly synthetic tasks; scores do not transfer to other decision problems. OOD here means the dataset's reserved wording/layout/family splits.
• Records longer than the max sequence length have the end of their state cut.
• Requires Unsloth's decision-model loader; the files are not a plain causal-LM adapter.