LFT_Final_FineTuned_Increased_Metrics

Model Card for LFT_Final_FineTuned_Increased_Metrics

Merged structure-aware LoRA deltas for the LfT math-tutoring student on top of Llama-3.1-8B-Instruct. This is the canonical LfT student for downstream math-tutor and IDC workflows.

Model Details

Model Description

• --Developed by:-- YRS Aakanksha
• --Shared by:-- YRS Aakanksha
• --Model type:-- Instruction-tuned causal LM with merged LoRA deltas (global LfT)
• --Language(s):-- English (math tutoring focus)
• --License:-- Same as base model (Llama-3.1-8B-Instruct)
• --Finetuned from:-- meta-llama/Llama-3.1-8B-Instruct

Model Sources

• --Repository:-- https://huggingface.co/Sashank-810/LFT_Final_FineTuned_Increased_Metrics

Uses

Direct Use

• Math tutoring / reasoning with structure-aware prompts (chapter/difficulty/LO tags).
• Base student for two-stage LfT + IDC flows.

Downstream Use

• Further task-specific fine-tuning for math reasoning or instructional tutoring.

Out-of-Scope Use

• Non-math domains; safety-critical decisions; any deployment without alignment/safety layers.

Bias, Risks, and Limitations

• Inherits biases and limitations of the base Llama-3.1-8B-Instruct model and the curated math datasets.
• Not safety-tuned; avoid use in safety-critical settings.

Recommendations

• Keep human oversight; add safety/filters for production.

How to Get Started

Load with transformers (merged weights)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Sashank-810/LFT_Final_FineTuned_Increased_Metrics"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

prompt = "Explain the concept of vector projections with an example."
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(--inputs, max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))

Serve with vLLM

pip install vllm
python -m vllm.entrypoints.openai.api_server \
  --model Sashank-810/LFT_Final_FineTuned_Increased_Metrics \
  --tensor-parallel-size 2 \
  --dtype auto

Then query via OpenAI-compatible endpoint (replace URL/key as needed):

import openai, os

openai.api_base = "http://localhost:8000/v1"
openai.api_key = "EMPTY"

resp = openai.ChatCompletion.create(
    model="Sashank-810/LFT_Final_FineTuned_Increased_Metrics",
    messages=[{"role": "user", "content": "Outline key steps in solving a probability problem involving Bayes' theorem."}],
    max_tokens=128,
)
print(resp["choices"][0]["message"]["content"])

Training Details

• Fine-tuned with structure-aware SFT across all chapters; LoRA deltas merged into base. Specific hyperparameters and dataset splits are kept private.

Evaluation

The model was evaluated on a comprehensive test set of 2,617 math tutoring questions, comparing performance against the base Llama-3.1-8B-Instruct model.

Accuracy Results

The fine-tuned model shows a --34.9% relative improvement-- in accuracy over the base model, with more than twice as many questions improved (421) compared to regressed (203).

Generation Quality Metrics

BLEU Score

The fine-tuned model achieves a --53.1% relative improvement-- in BLEU score (38.24 → 58.56), with significantly better length matching (BP: 0.612 → 0.993).

ROUGE Scores

All ROUGE metrics show substantial improvements, indicating better recall and overlap with reference answers.

METEOR Score

The METEOR score improvement demonstrates better semantic alignment and synonym matching in generated responses.

Key Findings

1. --Substantial Accuracy Gains--: The model demonstrates a clear improvement in mathematical correctness, with accuracy rising from 23.88% to 32.21%.

2. --Improved Response Quality--: Across all automated metrics (BLEU, ROUGE, METEOR), the fine-tuned model shows 27-53% relative improvements, indicating more coherent and relevant responses.

3. --Better Length Calibration--: The brevity penalty improvement (0.612 → 0.993) shows the model generates more appropriately-sized responses that better match expected answer lengths.

4. --Positive Net Impact--: With 421 improved questions versus 203 regressed, the model shows a strong positive impact ratio of approximately 2:1.

Technical Specifications

• Architecture: Llama-3.1-8B-Instruct with merged LoRA deltas (Phase 2 global LfT).
• Compute: Not disclosed; intended for GPU inference; vLLM compatible.

Model Card Contact

• Sashank-810 on Hugging Face