bhargavsai2005/flrrc-lora

FLRRC LoRA - Decision Engine (adapter v1.4)

Not a Chatbot. A Decision Calculator.

A small LoRA adapter on TinyLlama-1.1B-Chat that turns a personal decision question (English / Telugu-English mix) into a structured 5-box answer using the FLRRC framework: Fact, Logic, Reasoning, Representation (Equation + Graph), Conclusion.

Decision Support Tool only. The final decision is yours. This is a proof of concept: the output format is reliable, but the numbers and facts are not. See "Known limitations".

Model Details

• Developed by: Adhimulam Bhargav Sai Viswanath (bhargavsai2005)
• Model type: Causal LM + LoRA adapter (about 12.6 MB)
• Language(s): English, Telugu-English code-mixed
• License: MIT
• Finetuned from: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• Repository: https://github.com/AdhimulamBhargavSaiViswanath-05/flrrc-decision-engine
• Demo (Space): https://huggingface.co/spaces/bhargavsai2005/flrrc-decision-engine

Output format

The model writes compact JSON followed by the end-of-sequence token:

{
  "box_1_flrrc": {"fact": "", "logic": "", "reasoning": "", "representation": "", "conclusion": ""},
  "box_2_equation": {"formula": "", "go_score": "Go (..) = (v*w)+(v*w)-(v*w) = ... = x/10", "nogo_score": "", "regret_probability": ""},
  "box_3_graph": "Bar chart: Go X vs NoGo Y. note",
  "box_4_risk": {"best": "", "worst": "", "middle": ""},
  "box_5_next_action": ""
}

How to use

The prompt must match the training format exactly. There is no separate profile field: put the details in the question, in brackets.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
ADAPTER = "bhargavsai2005/flrrc-lora"

SYSTEM = (
    "You are FLRRC Decision Engine. Given a Question, output 5 boxes as JSON: "
    "box_1_flrrc, box_2_equation, box_3_graph, box_4_risk, box_5_next_action."
)

tokenizer = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.float32, device_map="cpu")
model = PeftModel.from_pretrained(model, ADAPTER).eval()

question = "Gym ki vellala? (energy 3/10, sleep 5 hrs, exam repu)"
prompt = f"<|system|>\n{SYSTEM}\n<|user|>\nQuestion: {question}\n<|assistant|>\n"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=800,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.05,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Recommended: do not trust the totals the model writes. Re-compute go_score / nogo_score from the (value*weight) terms in the string. The reference UI in the GitHub repo does this and shows a warning when the corrected winner differs from the model's own.

Training Details

Training data

75 hand-reviewed samples in ChatML format (data/training/chatml-v1.4.jsonl): 5 normalized gold samples plus 70 reviewed drafts. Go and NoGo totals in every sample are computed by code from the factors and weights. Winners: 37 Go / 38 NoGo. All fact, logic, reasoning and conclusion texts are unique. About 26 questions are English-style, the rest Telugu-English.

Training procedure

QLoRA on a Google Colab T4 (see notebooks/flrrc_v1_4.ipynb).

• Quantization: 4-bit nf4, fp16 compute
• LoRA: r=8, alpha=16, dropout=0.05, all attention and MLP projections
• Epochs: 5 (about 95 optimizer steps), batch 1, gradient accumulation 4
• Learning rate: 2e-4, cosine schedule, warmup 0.1
• Max length: 1024 tokens; </s> appended after each answer so the model learns to stop
• Loss: 2.52 (step 5) to about 0.95 (step 95)

Compared with v1.3

v1.3 trained for only 39 steps at the default learning rate on template-heavy data (loss 2.54 to 2.34) and did not produce the intended format. v1.4 fixes the data (no template filler), the learning rate, the step count and the end token.

Evaluation

No formal benchmark. Informal manual tests on a handful of questions (Colab and CPU):

• Valid 5-box JSON, stops at EOS: 4 of 4 in every informal run.
• Totals written by the model disagree with its own factors and weights in most samples checked (7 of 8 in one check).

Known limitations

• Arithmetic is unreliable. The formula is usually fine, the total often is not.
• Conclusion can contradict the scores. The text conclusion is free-form and may not match the numeric winner.
• Invents facts and numbers that are not in the question, and misreads vague or unusual questions.
• Trained on only 75 samples. Do not use for medical, legal or financial decisions.

Environmental Impact

• Hardware: NVIDIA T4 (Google Colab), about 15 minutes of training

Citation

@misc{flrrc-lora,
  author = {Adhimulam Bhargav Sai Viswanath},
  title = {FLRRC LoRA - Decision Engine (adapter v1.4)},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/bhargavsai2005/flrrc-lora}
}