vedatonuryilmaz/tpm-lite-qwen

TPM-Lite Qwen3

Parameter-efficient QLoRA fine-tune of Qwen/Qwen3-4B-Instruct-2507 for deterministic Technical Program Manager meeting extraction.

Target behavior: raw messy transcript in, strict Markdown ledger out. Output must start with # {MEETING TITLE} and include - **Source Fingerprint:** PENDING.

Artifacts planned in this repo:

• scripts/generate_tpm_data.py synthetic + real-leached data generator including lytang/MeetingBank-transcript
• scripts/train_qlora.py TRL SFTTrainer QLoRA script
• scripts/evaluate_schema.py strict schema/regex validator
• scripts/merge_and_export_gguf.sh LoRA merge and GGUF Q4_K_M export workflow
• LoRA adapter checkpoints and final merged/GGUF artifacts after training

Dataset repo: https://huggingface.co/datasets/vedatonuryilmaz/tpm-lite-data Base model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'vedatonuryilmaz/tpm-lite-qwen'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.