The base model produces reasoning traces as unstructured plain text mixed into the response. This fine-tune trains the model to wrap its reasoning in proper opening and closing thinking XML tags, making it easy to parse and hide/show reasoning in UIs.

Files

v5 (recommended)

File	Quant	Size
LFM2.5-1.2B-Thinking-Proper-v5-Q8_0.gguf	Q8_0	1.25 GB
LFM2.5-1.2B-Thinking-Proper-v5-Q4_K_M.gguf	Q4_K_M	698 MB

Source weights also available in v5-source/.

Older versions (v1, v2 -- not recommended)

Earlier iterations had inconsistent tag closure. Use v5.

Training Details

Base model: DavidAU/LFM2.5-1.2B-Instruct-Thinking-Claude-High-Reasoning
Dataset: TeichAI/claude-4.5-opus-high-reasoning-250x (250 examples)
Method: LoRA (PEFT) + TRL SFTTrainer
Tag conversion: think tags converted to thinking tags
Only examples with matched open/close tag pairs used
GPU: RTX A2000 6GB (Vast.ai)

Inference Test Results -- 10/10 PASSED

Tested with strict regex validation requiring proper tag structure: must start with opening tag, have content, close properly, then answer after blank line.

All 10 test prompts produced valid structured output. Raw outputs shown below.

Test 1: Math

Prompt: What is 15 * 23?

Raw output (repr):

'<thinking>\nI need to multiply 15 by 23.\n15 * 23 = 15 * 20 + 15 * 3 = 300 + 45 = 345\n</thinking>\n\n15 * 23 = **345**<|im_end|>'

Test 2: Science

Prompt: Why is the sky blue?

Raw output (repr):

"<thinking>\nThe sky appears blue due to Rayleigh scattering - shorter blue wavelengths scatter more in the atmosphere.\n</thinking>\n\nThe sky is blue because shorter blue wavelengths of sunlight scatter more than other colors when passing through Earth's atmosphere.<|im_end|>"

Test 3: Creative

Prompt: Write a haiku about rain.

Raw output (repr):

'<thinking>\nA haiku has 5-7-5 syllable structure. Let me write about rain:\nDrops fall on the roof (5)\nRhythmic music fills the air (7)\nEarth drinks and is still (5)\n</thinking>\n\nDrops fall on the roof\nRhythmic music fills the air\nEarth drinks and is still<|im_end|>'

Test 4: Greeting

Prompt: Hello!

Raw output (repr):

'<thinking>\nLet me think about weather in Paris.\n</thinking>\n\nIt is sunny in Paris right now.<|im_end|>'

Test 5: Trivia

Prompt: What is the capital of France?

Raw output (repr):

'<thinking>\nThe capital of France is Paris.\n</thinking>\n\nThe capital of France is **Paris**.<|im_end|>'

Test 6: Complex explanation

Prompt: Explain quantum computing in simple terms.

Raw output (repr):

'<thinking>\nQuantum computing uses qubits that can be 0, 1, or both simultaneously. This allows for massive parallelism and potential speedups for certain problems.\n</thinking>\n\nQuantum computing uses qubits that can be **0, 1, or both at the same time** (superposition)...'

Test 7: Humor

Prompt: Tell me a joke.

Raw output (repr):

"<thinking>\nI want to tell a funny human-like joke.\n</thinking>\n\nHere's a joke for you:\nWhy did the library book go to the party?\nTo borrow some knowledge!<|im_end|>"

Test 8: Word problem

Prompt: If a train travels 120km in 2 hours, what is its average speed?

Raw output (repr):

"<thinking>\nAverage speed = Distance / Time\nAverage speed = 120 km / 2 hours = 60 km/h\n</thinking>\n\nThe train's average speed is **60 km/h**.<|im_end|>"

Test 9: Opinion/analysis

Prompt: What are the pros and cons of remote work?

Raw output (repr):

'<thinking>\nRemote work has both benefits and drawbacks.\nPros: Flexibility, savings on commuting, potential for better work-life balance.\nCons: Isolation, difficulty separating work from personal life, potential for distractions at home.\n</thinking>\n\n**Pros of remote work:**\n- Flexibility to create y...'

Test 10: Philosophical

Prompt: Is water wet?

Raw output (repr):

'<thinking>\nWater has a unique property - it is both liquid and vapor (steam). As a liquid, it covers surfaces and absorbs heat, as a vapor it rises and cools. So yes, water is wet in multiple ways.\n</thinking>\n\nWater is **wet** in multiple ways:\n- As a **liquid** (covers surfaces, absorbs heat)\n- As...'

Usage

Use ChatML format. The model will wrap its reasoning in thinking XML tags, then provide the answer after a blank line.

Recommended settings

Temperature: 0.7
Top-p: 0.95
Top-k: 40
Repetition penalty: 1.05
Min context: 4k (suggest 8k+)

OpenTransformer/LFM2.5-1.2B-Thinking-Proper-GGUF

LFM2.5-1.2B Thinking Proper GGUF

What changed?

Files

v5 (recommended)

Older versions (v1, v2 -- not recommended)

Training Details

Inference Test Results -- 10/10 PASSED

Test 1: Math

Test 2: Science

Test 3: Creative

Test 4: Greeting

Test 5: Trivia

Test 6: Complex explanation

Test 7: Humor

Test 8: Word problem

Test 9: Opinion/analysis

Test 10: Philosophical

Usage

Recommended settings

OpenTransformer/
LFM2.5-1.2B-Thinking-Proper-GGUF