EvanHaacke's picture
Script tidy-up model, light export (q4, last-position logits)
2a3e200 verified
|
Raw History Blame Contribute Delete
1.27 kB
metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
library_name: transformers.js
pipeline_tag: text-generation
tags:
  - onnx
  - screenplay

Shot Planner script fix (0.5B)

A small model that repairs mistakes left after importing a screenplay PDF into Shot Planner (cue lines read as action, welded character names, name variants). It runs inside the app with Transformers.js, offline, and answers in the app's own JSON fix format; every answer is checked by the app before anything changes.

  • Base: Qwen2.5-0.5B-Instruct (Apache-2.0), LoRA fine-tune, merged.
  • Weights: onnx/model_q4.onnx (graph) + onnx/model_q4.onnx_data (weights): 4-bit layer weights, the tied embedding / output table in full precision, stored once, already transposed. Load with dtype: 'q4'; config.json tells Transformers.js to fetch the weights file (transformers.js_config.use_external_data_format).
  • Takes num_logits_to_keep (Transformers.js passes it on its own).
  • Same arithmetic as the earlier single-file export, laid out so the browser engine never builds a second copy of the table in its own memory.
  • Training data: synthetic screenplays and clean scripts damaged on purpose; the undamaged text is the answer.

Not a general chat model.