README.md

---
license: llama3.2
base_model: unsloth/Llama-3.2-1B-Instruct
tags:
  - llama
  - lora
  - qlora
  - gguf
  - health
  - biological-age
  - epigenetics
  - rag
language:
  - en
---

# epigenetics-slm

A Llama 3.2 1B Instruct model fine-tuned via QLoRA to generate five-category
epigenetic health assessments from wearable/biomarker data, grounded in a
Bio-RAG evidence retrieval pipeline.

Given a patient's biomarkers (HbA1c, NLR, circadian rest-activity metrics,
sleep architecture, CosinorAge acceleration) and retrieved evidence chunks,
the model produces a structured report with five sections: **AGING**,
**STRESS**, **METABOLISM**, **INFLAMMATION**, **SLEEP** — each citing the
evidence it was given.

## Files in this repo

| File | Description |
|---|---|
| `slm.q4_k_m.gguf` | Quantized model (q4_k_m), ~771MB, for CPU inference via `llama-cpp-python` |
| `slm_lora/` | Full LoRA adapter + tokenizer + all training checkpoints (100–1491 steps), with optimizer/scheduler state for resuming training |
| `chroma_db/` | Populated Bio-RAG vector store (47 chunks, `all-MiniLM-L6-v2` embeddings) — required alongside the model for grounded generation |

## Training details

- **Base model:** `unsloth/llama-3.2-1b-instruct-unsloth-bnb-4bit`
- **Method:** QLoRA via Unsloth + TRL `SFTTrainer`
- **Trainable params:** 11.2M / 1.25B (0.90%)
- **Dataset:** 4,415 examples derived from NHANES 2011-2012 biomarkers +
  Bio-RAG evidence (3,973 train / 442 eval)
- **Epochs:** 3 (1,491 steps), effective batch size 8
- **Hardware:** RTX 4070 Laptop GPU (8GB VRAM), WSL2 Ubuntu 22.04
- **Precision:** bf16 training, fp32 LoRA adapter weights

## Eval results (442-row held-out split)

| Metric | Score |
|---|---|
| Category coverage (all 5 headers present) | 99.8% |
| Classification match rate | 98.5% |
| ROUGE-L | 0.822 |

These metrics check structural adherence (all five sections present) and
classification-label accuracy against the training data's target
responses. They do **not** measure citation faithfulness — see
Limitations.

## Usage

Requires the model to be prompted via the exact training-time template
(see `slm_prompt.py` in the [source repo](https://github.com/rtsh13/epigenetics-slm)
for the canonical `build_prompt()`/`build_inference_prompt()` functions —
byte-identical prompt formatting between training and inference is
required for output quality).

```python
from llama_cpp import Llama

llm = Llama(model_path="slm.q4_k_m.gguf", n_ctx=4096, verbose=False)

# prompt must be built via build_prompt() + build_inference_prompt()
# from the source repo — see link above
out = llm(prompt, max_tokens=512, temperature=0.2, stop=["<|eot_id|>"])
print(out["choices"][0]["text"])
```

For full end-to-end usage (XGBoost baseline + Bio-RAG retrieval + this
SLM), see `scripts/demo.py` in the source repo.

## Limitations

- **Requires Bio-RAG evidence to ground citations.** When run without
  retrieved evidence chunks (empty RAG context), the model fabricates
  plausible-sounding but nonexistent citations (invented author names,
  journals, and DOIs). Always pass real retrieved evidence chunks.
- **Citation precision, not just presence.** Even with real evidence
  available, the model sometimes reuses the same citation pair across
  multiple report sections rather than mapping each specific claim to its
  most relevant source chunk. This is an attribution-precision issue, not
  fabrication.
- **1B parameter model.** Domain-specific acronym/definition accuracy
  (e.g. circadian rest-activity metrics) is not guaranteed to be robust;
  spot-check outputs before use in any downstream decision-making
  context.
- **Not a medical device.** This is a research/prototype system trained
  on NHANES survey data. Outputs should not be used for clinical
  decision-making.

## Source code

Training, export, and evaluation pipeline:
[github.com/rtsh13/epigenetics-slm](https://github.com/rtsh13/epigenetics-slm)
(branch `feat/week7-slm-finetuning`)