https://huggingface.co/firstbober/gemma-3-270M-it-smol-thinker

Here is an example of the output:
```
==================== THINKING ====================

Here is the thinking process:

- This is a large community with a wide range of interests

- Users can ask questions, share experiences, and discuss local events

- The rules are generally open-ended and allow for creativity

- However, the rules may be unclear or incomplete

<|thinking_end|>

==================== RESPONSE ====================

r/LocalLLaMA is a large, open-source question answering subreddit. Its rules are generally open-ended, allowing users to ask questions and share their experiences. However, the rules might be unclear or incomplete depending on the current state of the community.

<|response_end|>
```

It doesn't have much knowledge baked in, but with prompting it can give some interesting results.

Lore:

I've been working for a few days on it. First I just wanted to adapt it locally for function calling without using FunctionGemma. When it worked out (more or less) I moved to adding some thinking. The dataset was procedurally generated + some with Qwen 3.6 35B A3B (Q4 quants) + GLM 5.1.

The biggest hurdle was figuring out how to make it keep the format, I settled for rank 24, 768 max length for training data, and customized loss function which gives 20x for not using proper tags. Due to that the loss stayed at around 7, but the effect is there.

I've wanted to add longer examples, but my RTX 3050 4GB Mobile is kinda not enough, with train batch size of 1 and gradient accumulation step of 2 this is the best I could do.

Another interesting thing, Claude/Gemini were saying that bigger `gradient_accumulation_steps` essentially meant larger batch size but without actually increasing the batch size. This accounted for like 40% of all of my headaches, with model spitting utter garbage and random chinese slop characters.

Well, I think that's all, here are all the relevant training parameters:
```
SFTConfig:

per_device_train_batch_size=1,

gradient_accumulation_steps=2,

per_device_eval_batch_size=1,

learning_rate=1e-4,

lr_scheduler_type="cosine",

warmup_ratio=0.10,

weight_decay = 0.1,

load_best_model_at_end=True,

LoraConfig:

n_rank = 24

r=n_rank,

lora_alpha=n_rank,

target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],

lora_dropout=0.15,

task_type="CAUSAL_LM",
```

Oh, also increasing alpha to 2x rank as recommended in paper kinda broke everything, this is another thing that was pretty frustrating to figure out.

I plan to continue and train some more adapters with other ideas, maybe I'll switch to Qwen 3.5 0.8B when I buy a card with enough VRAM? I don't know. One thing I'll definitely do is thinking adapter for FunctionGemma, as it would fix my issues with function calling to some degree.