meeting-action-lora

Meeting Action Item Extractor (LoRA)

This repository contains a LoRA adapter fine-tuned to extract structured action items from meeting transcripts.

The model converts unstructured, multi-speaker conversations into strict, schema-valid JSON suitable for downstream automation, analytics, and workflow systems.

---

Task

Extract action items from meeting transcripts with the following schema:

{
  "action_items": [
    {
      "action": string,
      "owner": string | null,
      "deadline": string | null
    }
  ]
}

If no action items are present, the model outputs:

{ "action_items": [] }

---

Why This Project

Prompt-only extraction often:

• Misses implicit owners
• Hallucinates deadlines
• Breaks on meetings with no action items
• Produces inconsistent or non-parseable outputs

This project focuses on boringly reliable behavior instead of flashy demos:

• Strict output schema
• Explicit handling of edge cases
• Measurable evaluation

---

Model Details

• Base model: unsloth/mistral-7b-instruct-v0.2-bnb-4bit
• Fine-tuning method: LoRA (PEFT)
• Target modules: q_proj, v_proj
• Language: English
• Frameworks: Unsloth, PEFT, Transformers

This repository contains only the LoRA adapter, not the full base model.

---

Evaluation

• Metric: Schema validity (Pydantic validation)
• Dataset: Held-out real + synthetic meeting transcripts
• Result: 97.14% schema-valid outputs

For comparison, the base instruct model achieved near-zero schema validity due to format drift and verbose natural language outputs.

---

Training Data

• Manually labeled real meeting transcripts

• Synthetic meetings generated to cover:

  • Clear action items
  • Implicit ownership
  • Missing deadlines
  • Meetings with no action items

• Explicit negative examples to discourage hallucination

The dataset prioritizes format reliability and edge-case coverage, not scale.

---

Usage

from unsloth import FastLanguageModel
import torch

# Load base model
model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
    load_in_4bit=True,
    max_seq_length=2048,
)

# Attach LoRA adapter
model = FastLanguageModel.get_peft_model(
    model,
    r=8,
    lora_alpha=16,
    target_modules=["q_proj", "v_proj"],
)

model.load_adapter(
    "muskankh03/meeting-action-item-lora",
    adapter_name="default"
)

---

Limitations

• Designed for English business meetings
• Does not infer missing information beyond the transcript
• Not intended for summarization or open-ended chat

---

License

Apache 2.0 (adapter only; base model license applies separately)

---

Author

Muskan Khandelwal