Qwen3.5-2B ยท Unsloth Decision head (Clef) โ€” project-browser, Q4_K_M

A decision model, not a chat model. Given a state (task + step + previous actions) and a set of questions whose options are page elements, it returns the option key to interact with next, plus a calibrated probability per option.

Served by Unsloth Studio's /v1/systemone (llama.cpp backend, clef layout).

Files

file size note
model-Q4_K_M.gguf 1.29 GB full checkpoint: backbone + joint decision head
mmproj-Q8_0.gguf 365 MB vision tower (only needed for image input)
export.json โ€” Studio export contract (layout, quantization, fingerprint)
unsloth_decision_config.json โ€” temperatures, training metadata
joint_head_config.json โ€” head shape (hidden 2048, width 512, 4 layers)

Do not run this with a chat prompt. It needs a state + questions request.

Request shape

POST /v1/systemone
{
  "model": "default",
  "state": "Task: click the Interfaces row in the Health table\nWebsite: example.com\nStep: 4 of 8\nPrevious actions:\n(none)",
  "questions": {
    "target": {
      "type": "choice",
      "instructions": "Which element on the page should be interacted with next ...",
      "criteria": {"e00": "<div role=\"row\"> Interfaces 86 50 1", "e01": "...", "none": "None of the listed elements ..."}
    },
    "operation": {"type": "choice", "instructions": "What should be done ...",
                  "criteria": {"click": "CLICK ...", "type": "TYPE ...", "select": "SELECT ..."}},
    "needs_value": {"type": "noul", "instructions": "Does this action need a specific value ..."}
  }
}

Response: answers.target.choice + confidence + probabilities (calibrated).

Envelope (measured, do not exceed)

  • โ‰ค10 options per question (+ none). The calibration temperatures are fit per option-count bucket (โ‰ค2 / 3โ€“5 / 6โ€“10 / 11+); at 41โ€“53 options the distribution becomes uninformative (flat or sharp-but-wrong).
  • Atomic instruction (one step, English). Multi-step SOPs are the planner's job, not this head's.
  • Website: = hostname only. A full URL leaks tenant/slug tokens that anchor the head to unrelated elements.

Evaluation (eval split, N=375, target accuracy)

quantization accuracy ECE
Q8_0 / MLX fp16 (source) 88.3% 0.0506
Q4_K_M (this file) 86.9% 0.0422

Agreement Q8 vs Q4: 96.3%. none recall 100% in both; no false abstentions. The joint head is quantized too (Q4_K/Q6_K weights, F32 norms/bias) โ€” the measured cost is 1.4 pp accuracy with slightly better calibration.

Intended use / limits

For: picking the next element in a browser agent loop, one step at a time, with a pre-filtered candidate list (a ranker or a planner supplies โ‰ค10 options).

Not for: free-text generation, writing field values, judging whether a task is complete, planning multi-step procedures, or searching a large DOM (it does not rank hundreds of elements โ€” pass it a shortlist).

Provenance

  • Base: unsloth/Qwen3.5-2B (QLoRA, soft_cross_entropy, clef head).
  • SFT data: Mind2Web-derived decision rows (train/eval splits, ~6.9k steps). Mind2Web is CC-BY-4.0; derivative data keeps that attribution.
  • Exported from an Unsloth Studio decision checkpoint; see export.json source_fingerprint for the exact source folder state.
Downloads last month
110
GGUF
Model size
2B params
Architecture
clef
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for nerozx/qwen3.5-2b-decision-browser-q4

Finetuned
Qwen/Qwen3.5-2B
Quantized
(22)
this model