rcrc-chat-v5-gemma-1b-cpt-sft/README.md

---
license: gemma
base_model: perfectPresentation/rcrc-gemma-1b-cpt
datasets:
- perfectPresentation/rcrc-qa-v5
language:
- ar
- en
tags:
- chat
- rcrc
- arabic
- gemma
- gemma3
- closed-book
library_name: transformers
---

# rcrc-chat-v5-gemma-1b-cpt-sft — Closed-Book RCRC Chatbot

A 1B closed-book chatbot for the Royal Commission for Riyadh City (RCRC).
Trained as **Path B** of the v5 comparison: continued pre-training on raw KB
text, then chat SFT on Qwen3-235B-synthesized QA pairs.

Closed-book here means the model answers from baked-in knowledge — there is
no retrieval at inference.

## Pipeline

```
google/gemma-3-1b-pt
   ↓  CPT: 3 epochs on cleaned RCRC + Hanifa raw text
perfectPresentation/rcrc-gemma-1b-cpt
   ↓  SFT: 3 epochs on rcrc-qa-v5 (16,761 Qwen-synthesized QA pairs)
perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft   ← this repo
```

## Training data

[`perfectPresentation/rcrc-qa-v5`](https://huggingface.co/datasets/perfectPresentation/rcrc-qa-v5):
16,426 train + 335 validation single-turn (system, user, assistant) pairs
synthesized by `Qwen/Qwen3-235B-A22B-Instruct-2507` from the cleaned RCRC
website + Hanifa Urban Code chunks.

## Training recipe

| | |
|---|---|
| Base | `perfectPresentation/rcrc-gemma-1b-cpt` |
| Epochs | 3 |
| LR | 2e-5, cosine, 5% warmup |
| Effective batch | 16 (per_device 4 × grad_accum 4) |
| Max seq length | 1024, packing enabled |
| Final eval loss | 0.71 |
| Final eval token-accuracy | 83.5% |
| Hardware | HF Jobs · 1× L4 (~1.6 h) |

## Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft")
model = AutoModelForCausalLM.from_pretrained(
    "perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft",
    dtype=torch.bfloat16,
)

messages = [
    {"role": "system", "content":
        "أنت مساعد للهيئة الملكية لمدينة الرياض. تجيب على استفسارات المستخدمين "
        "عن خدمات وبرامج ومشاريع وأنظمة الهيئة بدقة وأدب."},
    {"role": "user", "content": "ما هو الكود العمراني لوادي حنيفة؟"},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs, max_new_tokens=400, do_sample=True, temperature=0.5, top_p=0.9,
    repetition_penalty=1.15, no_repeat_ngram_size=6,
)
print(tok.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
```

## Honest evaluation: closed-book vs RAG

On a 50-question internal eval set (RCRC + Hanifa + edge cases), this
closed-book model was compared to a sibling RAG pipeline (the v3-rag
checkpoint reading retrieved chunks from `rcrc-rag-index-v2`):

| | accuracy | relevance | clarity | wins |
|---|---:|---:|---:|---:|
| RAG (v3-rag + index v2) | 3.92 | 4.08 | 4.42 | 34/50 |
| **This model (closed-book)** | **2.64** | **3.36** | **3.56** | **13/50** |
| Ties | | | | 3/50 |

**Where closed-book is competitive**: dialect responses (Najdi/Hijazi),
free-form opinion-style queries.

**Where RAG dominates**: factual specifics from the Hanifa Urban Code,
project details, organizational facts, numerical specs.

For accuracy-sensitive deployments on the RCRC + Hanifa corpus, a RAG
pipeline at the same parameter count outperforms this closed-book model
by ~1.3 points on average. This model is provided for completeness of the
v5 study and for offline / no-retrieval scenarios.

## Limitations

- 1B-scale closed-book recall is brittle on specifics (numbers,
  exact procedure steps). Verify against rcrc.gov.sa.
- Training data was Qwen-synthesized; some questions may carry the
  synthesizer's biases.
- No multi-turn conversational SFT — single-turn QA only.

<!-- GGUF-BUILDS -->
## GGUF builds (llama.cpp / Ollama)

Quantized GGUF files live at the repo root.

| File | Quant | Approx size |
|---|---|---|
| `rcrc-v5-gemma-1b-cpt-sft-F16.gguf` | F16 | ~2.0 GB |
| `rcrc-v5-gemma-1b-cpt-sft-Q8_0.gguf` | Q8_0 | ~1.0 GB |
| `rcrc-v5-gemma-1b-cpt-sft-Q5_K_M.gguf` | Q5_K_M | ~720 MB |
| `rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf` | Q4_K_M | ~620 MB |

### Ollama (one-liner)

```bash
ollama run huggingface.co/perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft:Q4_K_M
# or :Q8_0, :Q5_K_M, :F16
```

If you hit a host-redirect error (`hf.co` → `huggingface.co`), upgrade
Ollama to a recent version, or use the `huggingface.co/...` URL above.

### Manual Modelfile route

```bash
hf download perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf Modelfile --local-dir ./model
cd model
ollama create rcrc-v5-gemma-1b-cpt-sft -f Modelfile
ollama run rcrc-v5-gemma-1b-cpt-sft
```

### llama.cpp

```bash
hf download perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf --local-dir .
./llama-cli -m rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf -cnv
```

## GGUF builds (llama.cpp / Ollama)

Quantized GGUF files live under `gguf/`. They are built directly from
the safetensors above with `llama.cpp convert_hf_to_gguf.py` followed by
`llama-quantize`.

| File | Quant | Approx size |
|---|---|---|
| `gguf/rcrc-v5-gemma-1b-cpt-sft-F16.gguf` | F16 | ~2.0 GB |
| `gguf/rcrc-v5-gemma-1b-cpt-sft-Q8_0.gguf` | Q8_0 | ~1.0 GB |
| `gguf/rcrc-v5-gemma-1b-cpt-sft-Q5_K_M.gguf` | Q5_K_M | ~720 MB |
| `gguf/rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf` | Q4_K_M | ~620 MB |

### llama.cpp

```bash
hf download perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft gguf/rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf --local-dir .
./llama-cli -m gguf/rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf -cnv
```

### Ollama

```bash
hf download perfectPresentation/rcrc-chat-v5-gemma-1b-cpt-sft gguf/rcrc-v5-gemma-1b-cpt-sft-Q4_K_M.gguf gguf/Modelfile --local-dir ./model
cd model/gguf
ollama create rcrc-v5-gemma-1b-cpt-sft -f Modelfile
ollama run rcrc-v5-gemma-1b-cpt-sft
```