Llama-3.2-3B-ReceivablesChaser-GGUF

Llama-3.2-3B-ReceivablesChaser-GGUF

A LoRA fine-tune of Llama 3.2 3B Instruct, trained to act as the offline judgment layer for Receivables Chaser — a hybrid rules-and-AI agent for B2B receivables recovery, built for the Razorpay AI Buildathon (Track 3: AI Revenue Recovery).

Full project, source code, and architecture writeup: github.com/shaan404/receivables-chaser

What this model does

Given a B2B invoice's context (amount, days overdue, payment history, relationship tier, and free-text notes) that a separate, non-negotiable rules engine has already cleared for automated contact, this model:

1. Drafts a personalized outreach message matched to the invoice's ladder stage (friendly reminder / second reminder / formal notice)
2. Flags the invoice for human review instead, if something in the notes looks unusual — a dispute, a bankruptcy mention, a mismatch between structured data and what the notes actually say

It never decides whether an invoice can be contacted — that stays entirely with the deterministic rules engine, outside this model entirely.

Training details

• Base model: unsloth/Llama-3.2-3B-Instruct-bnb-4bit
• Method: LoRA (r=16, alpha=16) via Unsloth, training under 1% of the base model's parameters
• Data: 500 examples, distilled from Claude (claude-haiku-4-5) labeling synthetic invoices with the exact same system prompt used in production
• Hardware: free Google Colab T4 GPU, ~19 minutes training time
• Quantization: q4_k_m GGUF, for fast CPU-only inference via llama-cpp-python

The full, runnable training notebook is in the main repo: docs/finetune_colab_notebook.ipynb

Usage

from llama_cpp import Llama

llm = Llama(model_path="Llama-3.2-3B-Instruct.Q4_K_M.gguf", n_ctx=2048)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "<see llm_agent.py in the main repo for the exact system prompt>"},
        {"role": "user", "content": "<invoice context — see local_agent.py for the exact format>"},
    ],
    max_tokens=300,
    temperature=0.7,
)
print(response["choices"][0]["message"]["content"])

For the full working integration (prompt construction, JSON parsing, error handling), see local_agent.py in the main repo.

License

Built with Llama. Use of this model is governed by the Llama 3.2 Community License. This is a derivative of Meta's Llama 3.2 3B Instruct model.

A note on limitations

Spot-testing against the base Claude model (used to generate this model's training data) showed this fine-tune is more conservative than Claude on genuinely ambiguous edge cases — e.g., a case involving a mismatch between structured contact-history data and free-text notes, where Claude itself was inconsistent across repeated runs (2 of 3 flagged), and this model consistently did not flag it (0 of 3). This is disclosed here rather than hidden — the model is not a drop-in equivalent for Claude on every case, particularly ambiguous ones.