phi4RAG

Phi-4 RAG (LoRA fine-tuned) — Q4_K_M GGUF

Quantized GGUF build of microsoft/phi-4 with a LoRA adapter merged in, fine-tuned for retrieval-augmented question answering. The model answers only from supplied document context in English, Spanish, or Catalan, using the same RAG-oriented system prompt as MonkeyGrab, a local, fully private RAG stack developed for a Bachelor's thesis (TFG) at the Universitat Politècnica de València (UPV).

Source code, thesis, and contact

The full MonkeyGrab source code is publicly available at:

https://github.com/iDiagoValeta/localOllamaRAG

The repository includes the complete RAG pipeline, CLI, web interface, training scripts, evaluation workflows, and documentation for the Bachelor's thesis (TFG) at UPV.

This Hugging Face model repo ships inference assets (Phi4-Q4_K_M.gguf), the Ollama Modelfile, and a reproduction/ folder with frozen copies of the training script, merge utility, and evaluation_comparison.json so methodology and metrics remain auditable alongside the full codebase.

Contact: [email protected] for questions about training, evaluation, or Ollama usage.

GGUF pipeline (high level): LoRA fine-tuning on the datasets below → merge with merge_lora.py (see reproduction/) → GGUF export via the llama.cpp toolchain → Q4_K_M quantization. The merge script documents expected paths and flags.

Files in this repo

Base model and method

• Base: microsoft/phi-4 — 14B-parameter transformer (ChatML-style; end-of-turn token <|im_end|>).
• Adaptation: PEFT LoRA fine-tuning on five RAG-focused datasets → LoRA adapter merged into dense weights → GGUF export → Q4_K_M quantization.

LoRA configuration

Training (train-phi4.py, v1)

• Seed: 42 (propagates to torch / NumPy / CUDA via set_seed).
• Task format: ChatML <|im_start|>user … <|im_end|> with the instruction and <context>…</context> on the user turn; loss computed only on the assistant completion (prompt labels masked with –100).
• Data — balanced 5-way interleaving (3,200 train samples per source, 16,000 total):
  • neural-bridge/rag-dataset-12000
  • databricks/databricks-dolly-15k (categories: closed_qa, information_extraction, summarization) — 80/10/10 split after filter
  • projecte-aina/RAG_Multilingual — EN, ES, CA subsets
• Sequence limits: max_length 4,096 tokens; context truncated to 2,048 tokens; generation up to 2,048 new tokens.
• Optimizer / schedule: AdamW 8-bit, lr 5e-5, cosine decay with warmup_ratio 0.05, weight_decay 0.01, max_grad_norm 1.0.
• Batching: per_device_train_batch_size 1, gradient_accumulation_steps 16 → effective batch 16; bf16 + TF32; gradient checkpointing enabled.
• Epochs: 3; checkpoints saved every 300 steps (keep last 3); eval every 150 steps; load_best_model_at_end on eval_loss; early stopping patience 3 evaluations.

Evaluation protocol

• Frozen dev/test splits: identical for the base (microsoft/phi-4) and the adapted (LoRA merged) model — no data leakage.
• Dev: 320 samples × 5 sources = 1,600 examples (aligned with evaluate_baselines.py for cross-experiment comparability).
• Test: full held-out splits — 8,490 examples total across all five sources.
• Metrics: Token F1, ROUGE-L F1, BERTScore F1 (microsoft/deberta-xlarge-mnli); BERTScore is computed after unloading the generative model to fit in GPU memory.
• Artifacts: all metric values and sample pairs are in reproduction/evaluation_comparison.json.

Evaluation results

Values are percentage points (0–100 scale). Δ (pp) = adapted − base; Δ rel (%) = relative change vs. base.

Weighted aggregate (all five sources)

Per-dataset dev (320 samples each)

Full test-split breakdowns and qualitative sample pairs are in reproduction/evaluation_comparison.json.

Relation to the baseline benchmark

The base dev numbers are aligned with the multi-model benchmark (evaluate_baselines.py, predictions_phi-4.json), so Phi-4 before fine-tuning is directly comparable to the other models in that suite. For post-LoRA performance, use the Adapted columns above.

Hardware compatibility (inference)

Ollama

Place Phi4-Q4_K_M.gguf next to Modelfile (or adjust the FROM path). Then:

ollama create phi4-rag -f Modelfile
ollama run phi4-rag

Generation defaults in the bundled Modelfile: num_ctx 16384, temperature 0.15, top_p 0.9, repeat_penalty 1.15.

Limitations

• Intended for grounded QA over retrieved context; do not rely on it as an unconstrained world-knowledge model without retrieval.
• Q4_K_M is a speed/size trade-off versus higher bit-widths or FP16.
• Response quality depends on the quality of the retrieved context and on wrapping it in <context>…</context> tags as in training.

License

• MIT — The model card, Modelfile, and other metadata added by nadiva1243 are released under the MIT License (see the LICENSE file in this repository).
• Base weights — The GGUF is a derivative of microsoft/phi-4. You must also comply with the license and terms of the base model and with any requirements of the training datasets when redistributing or using the weights.

Citation

@misc{phi4_rag_gguf_monkeygrab,
  title        = {Phi-4 RAG LoRA Fine-tune (Q4_K_M GGUF)},
  author       = {nadiva1243},
  year         = {2026},
  howpublished = {Hugging Face: \url{https://huggingface.co/nadiva1243/phi4RAG}},
  note         = {Base: microsoft/phi-4; training: MonkeyGrab train-phi4.py v1; source: https://github.com/iDiagoValeta/localOllamaRAG}
}