---

license: apache-2.0

language: [en]

tags: [auroragpt, gguf, reasoning, hybrid-thinking, on-device, small-language-model]

---

AuroraGPT-700M-Reason (GGUF, Q4_0)

Hybrid-reasoning SFT of the raw AuroraGPT-700M base on

openbmb/UltraData-SFT-2605 -- the corpus MiniCPM5-1B was trained on --

followed by a light identity/greeting pass. By UltraLabs.

• 707.5M params, Llama-style, context 4096 (2x extension, rope_theta=20200)
• Q4_0, 412 MB, EOS <|end|> (id 5)
• Reasoning SFT: 40k rows, 50/50 think/no-think, 2 epochs, 103.8M tokens

(loss 1.53 -> 0.86). Then 29k rows at 1e-5 for identity and casual register,

82.8% of it rehearsal so the reasoning is preserved.

Thinking toggle

The model emits <think> on its own, like Qwen3. The template only intervenes

to disable it:

| mode | generation prompt |

|---|---|

| think (default) | <\|assistant\|> |

| no-think | `<\|assistant\|><think>

</think>

` |

Set enable_thinking=false for direct answers. The chat template round-trips

reasoning_content, so llama.cpp's autoparser reports

reasoning_mode: TAG_BASED and clients such as PocketPal show a think toggle

automatically.

12% of training traces were cut mid-thought, so a client-side thinking budget

(force-injecting </think> at a token cap) degrades gracefully.

What it does well

Sustains long reasoning traces on-topic -- multi-thousand-token chains that

stay coherent, and it will catch and correct itself mid-trace rather than

committing to a wrong first answer. Knows its identity and handles casual

conversation.

Honest limitations

Arithmetic is unreliable. Sampled repeatedly, 17 + 25 returned 38, 46, 48,

85 and once a correct 42, each with a fluent derivation. The base saw 22.4B

pretraining tokens (~32 tokens/param, versus ~14,800 for SmolLM2-135M), so

arithmetic never emerged and SFT cannot install it. Use a calculator tool.

No safety tuning. Refusal data was tried and removed: at this scale it

learned "topic sounds dangerous -> decline" and refused legitimate questions

(occupational medicine, defensive security) while engaging with genuinely

harmful ones. Filtering belongs outside the weights for a model this size.

Built on SmallAICreator/AuroraGPT-700M.