jiwo¶
| Checkpoints | Inputs |
|---|---|
eljiwo/jiwo-0.8b, -4b |
text |
jiwo-0.8b is a decision model on Qwen/Qwen3.5-0.8B and jiwo-4b one on Qwen/Qwen3.5-4B, each with a readout head in
readout.safetensors. Flide asks them with the release's own prompt. Each question is a chat prompt of its own. The
system turn asks for the code of the selected option and tells the model to treat the state as data. The user turn
holds the state, a heading that depends on the question's type, and the options, and asks for the code of the best
option. A state that is not a string is written as JSON. A noul is headed "Statement (decide whether it is true for the
state)" and lists "No, the statement is false." and then "Yes, the statement is true." unless the request describes the
two outcomes. A score is headed "Question (a rating scale, lowest level first)" and lists its levels. A choice is headed
"Question" and lists each option's name and, when it has one, its description. An option name that looks like a code,
one or two capital letters or a number of up to three digits, is written as option "NAME".
The options are labelled with the codes listed as codes in decision_config.json. The readout head has one row per
code, and the answer is the softmax over the rows of the question's options at the last position of the prompt. The
temperature of each question type comes from type_temperatures in decision_config.json. The state comes first in
every prompt, so Flide computes it once and forks it into the questions. Thinking is off.
Speed¶
On one H200, one request at a time, Flide answers 1.7 to 6.9 times faster than jiwo's own engine with its CUDA graphs. Each workload is 32 requests, each run once on both sides before timing, and the table gives the median latency of a request. MultiRC requests ask about 49 questions each.
| Checkpoint | Workload | Flide | Release | Speedup |
|---|---|---|---|---|
| jiwo-0.8b | Decision Index | 4.8 ms | 10.0 ms | 2.1× |
| jiwo-0.8b | RACE | 3.9 ms | 8.1 ms | 2.0× |
| jiwo-0.8b | MultiRC | 27.8 ms | 137.5 ms | 5.0× |
| jiwo-0.8b | QuALITY long | 22.0 ms | 39.2 ms | 1.8× |
| jiwo-4b | Decision Index | 13.2 ms | 23.0 ms | 1.7× |
| jiwo-4b | RACE | 8.9 ms | 15.6 ms | 1.7× |
| jiwo-4b | MultiRC | 75.0 ms | 518.1 ms | 6.9× |
| jiwo-4b | QuALITY long | 66.5 ms | 116.5 ms | 1.8× |
Start a server¶
flide serve --model eljiwo/jiwo-4b
Send a request¶
curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
"state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
"questions": {
"action": {"type": "choice", "instructions": "What should support do next?",
"criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
"urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
}
}'
Limits and options¶
A request takes choice, noul and score questions. A choice takes as many options as the checkpoint lists codes, up to 255, and a score takes as many levels. The longest prompt is 32768 tokens, and the state is not cut; a longer prompt is refused. A request with images is refused. The template refuses nothing else.
| Option | What it does |
|---|---|
--format template --format-arg spec=eljiwo__jiwo |
reads a copy of the checkpoint kept in a directory Flide does not recognize |