---
base_model: schneewolflabs/A0i-12B
datasets:
- schneewolflabs/BigDenker-SFT
library_name: transformers
pipeline_tag: text-generation
tags:
- reasoning
- thinking
- chain-of-thought
- mistral
- sft
- lora
license: apache-2.0
---
# A1
A1 is a reasoning-tuned version of [`schneewolflabs/A0i-12B`](https://huggingface.co/schneewolflabs/A0i-12B) (Mistral Nemo–class, 12B). It was supervised-fine-tuned on [`schneewolflabs/BigDenker-SFT`](https://huggingface.co/datasets/schneewolflabs/BigDenker-SFT) to produce explicit `…` chain-of-thought before its final answer.
## What's different from the base model
The base A0i-12B does not reason — given a prompt it answers directly. A1 produces a reasoning trace inside `` and then the answer, in the Qwen3 thinking convention.
The reasoning tokens were added **without resizing the vocabulary**. A0i's tokenizer ships with 986 unused reserved slots (``…``); ten of these were repurposed in place (token IDs unchanged, so embeddings were *not* resized):
| ID | token | ID | token |
|----|-------|----|-------|
| 14 | `` | 19 | `` |
| 15 | `` | 20 | `<\|vision_start\|>` |
| 16 | `` | 21 | `<\|vision_end\|>` |
| 17 | `` | 22 | `<\|image_pad\|>` |
| 18 | `` | 23 | `<\|video_pad\|>` |
Before training, these rows (which were zero/untrained in the base) were initialized from the mean of their surface string's sub-token embeddings (computed separately for `embed_tokens` and `lm_head`, which are untied), with a small symmetry-breaking perturbation. They were then trained as part of the finetune.
> Note: the vision/tool tokens exist for chat-template completeness. A1 is a **text-only** model and was not trained on vision or tool-calling data.
## Usage
A1 uses a Qwen3-style chat template (bundled as `chat_template.jinja`). The template injects `\n` as the assistant prefix, so the model continues with its reasoning, emits ``, then the answer. **Always use the chat template** — using the model without it will not trigger reasoning.
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("schneewolflabs/A1")
model = AutoModelForCausalLM.from_pretrained(
"schneewolflabs/A1", dtype=torch.bfloat16, device_map="auto"
)
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "A bat and ball cost $1.10. The bat is $1.00 more than the ball. How much is the ball?"},
]
enc = tok.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(**enc, max_new_tokens=1024, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=False))
```
The output has the form `…reasoning…\n\n…final answer…<|im_end|>`.
## Training
- **Method:** SFT, 1 epoch, via the Merlina training system (grimoire `SFTLoss`, prompt-masked completion).
- **Adaptation:** LoRA (r=64, α=128, dropout 0.05) on attention + MLP projections, plus `embed_tokens` and `lm_head` as fully-trained `modules_to_save` (required so the repurposed token rows actually learn).
- **Hyperparameters:** lr 2e-5 (cosine, 5% warmup), effective batch 16 (bs 1 × grad-accum 16), max sequence length 4096, bf16, seed 42.
- **Split:** 90% train / 10% held-out eval (random, seed 42).
- **Result:** train loss 1.22 → 0.687; held-out eval loss ≈ 0.653 (eval ≤ train — no overfitting at 1 epoch).
- The conservative learning rate plus the semantic embedding initialization were chosen to add reasoning while limiting drift of the base model's general token representations (the full embedding/lm_head matrices were trainable).
## Evaluation notes
Behavioral checks show coherent step-by-step reasoning that **generalizes beyond the training distribution** — e.g. it solves the bat-and-ball problem correctly ($0.05) and explicitly rejects the common intuitive-trap answer.
## Limitations
- **Always-on thinking:** the template starts every assistant turn with ``; the model reasons even on trivial prompts. A non-thinking path exists via the template (`enable_thinking=False` injects an empty ``) but was not specifically tuned.
- **Single-source SFT:** trained on one dataset/style (BigDenker), so reasoning phrasing is fairly homogeneous.
- **One epoch / conservative LR:** a deliberate, safe first pass — not exhaustively tuned.
- Inherits all limitations and biases of the base model and the SFT data.
## Provenance
Base: `schneewolflabs/A0i-12B` · Data: `schneewolflabs/BigDenker-SFT` · Tokenizer/template: repurposed reserved tokens + Qwen3 chat template.