donotfail/insurance-CTP-DSLM

This GGUF has been fine-tuned using Qwen3 8B. Trained for NRMA CTP Insurance Policy for ACT state in Australia using Jupyter Notebook on Google Colab T4 NVIDIA. It has been shrunk from 16 GB to 5 GB Q4_K_M 8B. Tools used to fine-tune are from unsloth and Hugging Face.

The Model can run on a general laptop CPU and RAM.

#running the insurance model

Download the insurance model (.gguf), the NRMA booklet (.pdf) and role.txt to your computer. The model is 5 GB Q4_K_M 8B, so you need roughly 8 GB of free RAM or VRAM.

Then run the model in LM Studio (easiest) or Ollama. Ask Claude which folder to move the .gguf file to, depending on whether you are using LM Studio or Ollama. You may need to change the default settings of those tools to CPU if you don't have GPU.

Before chatting, give the insurance model some instructions on its role (cut & paste prompt in role.txt into chat). Then upload the NRMA booklet .pdf to the chat (poor man's RAG).

You are ready to chat. Try "I have fractured my pelvis. Can I make a claim?"

#limitations

What is still being built around this model is conversational memories (1 & 2 are already built into the Model):

1. Working Memory: YES. DONE. Model already uses the context window and KV cache to process the current conversation.

2. Parametric Memory: YES. DONE. Knowledge and learned patterns encoded in the model's weights during training and fine-tuning.

3. Semantic Memory: PARTIAL. TBD. Model has learned general facts in its weights, but doesn't automatically save new user preferences across sessions.

4. Procedural Memory: PARTIAL. TBD. Model can follow instructions in context (role.txt has been used to override some unwanted instructions); persistent custom workflows require additional storage or tooling.

5. Episodic Memory: NO. TBD. Needs a system to store and retrieve past interactions and outcomes.

6. Retrieval Memory: NO. TBD. For now uploading fact based PDF into chat context is enough. Really requires a retrieval system such as RAG connected to documents or a vector database.

7. Prospective Memory: NO. TBD. Requires a scheduler, task database, or reminder system to remember and execute future tasks.

#"holy grail"

1. The Judgement Layer: Code currently in testing. Combines domain ontology, domain rules, evidence, evaluate situations, and execute validated actions deterministically. The code for this will be released on GitHub soon.

THIS IS MY PERSONAL WORK BASED ON OPENSOURCE STUFF. USE IT ON "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND. Or something along those lines from Apache-2.0.