Aureliolo/steamgauge-writer

SteamGauge writer

Qwen/Qwen3-4B-Instruct-2507 as an ONNX graph with int4 weights, rounded to nearest in blocks of 32 by onnxruntime-genai's model builder. SteamGauge uses it to write the paragraph at the top of a game's page from the points its Steam reviewers make most, and leaves out any sentence giving a number or a quotation those points do not give.

Use

• model.onnx takes input_ids from tokenizer.json, attention_mask, and past_key_values.{layer}.key and .value for each of its 36 layers, in single precision, 8 heads of 128; it returns logits and present.{layer}.key and .value, which are the next step's past. The weights are in model.onnx.data.
• Input is the model's chat format. SteamGauge takes the likeliest token at every step and stops at <|im_end|> or <|endoftext|>.
• It is built for the processor's int4 kernels and runs on the processor. On DirectML it runs slower and its answers are not the model's.
• training/export_writer.py builds it and checks it against the original model: fed the original's own paragraph for a fixed brief, it has to pick the original's next token at three places in four or more.

Licence

Apache-2.0, as is Qwen/Qwen3-4B-Instruct-2507. The model is Qwen's work.