Qwen3.5-2B ยท Unsloth Decision head (Clef) โ€” project-browser, Q3_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-Q3_K_M.gguf 1.11 GB full checkpoint: backbone + joint decision head
export.json โ€” export contract for Q3_K_M
unsloth_decision_config.json โ€” temperatures, training metadata
joint_head_config.json โ€” head shape (hidden 2048, width 512, 4 layers)

The vision tower (mmproj) is not included in this repo; use the Q4 repo's mmproj-Q8_0.gguf if image input is required.

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

Quantization

  • Quantized from the original F16 GGUF (model-F16.gguf), not from Q4.
  • Size: 1,109,253,184 bytes vs Q4_K_M 1,293,525,216 bytes (14.3% smaller).
  • SHA-256: 48e22f2ee366e6ec77386bae666ffdb4c5d4cddba239c01c35791bb511b57084

Evaluation

Test Result
Training heldout (400 decisions, Q4-era checkpoint) 96.5% accuracy, ECE 0.0152
100-row b297, primitive head-only 93% accuracy, ECE 0.0640, false-accept 7
Real-goal NMS/Grafana (42 goals, 3 reps) 69.4% positive accuracy, 30.6% wrong, 0% abstain

The 96.5% figure is the training heldout, not a final out-of-distribution eval. The real-goal benchmark measures the full decision tool (snapshot + embedding + sequential decide), not just the head.

Usage

llama-server \
  --model model-Q3_K_M.gguf \
  --port 8899 \
  -ngl 99 -b 4096 -ub 4096

The server must expose the clef decision layout (/v1/systemone). For the browser tool, set:

export DECISION_BASE=http://127.0.0.1:8899
export DECISION_MODEL=model-Q3_K_M
export DECISION_BACKEND=unsloth

Tool + benchmark: nifira-agent/decision-tools (Gitea), branch feat/sequential-embedding.

License

Other; follow the base model and training-data licenses.

Downloads last month
63
GGUF
Model size
2B params
Architecture
clef
Hardware compatibility
Log In to add your hardware

3-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-q3

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