README.md

---
base_model: schneewolflabs/A1.1
datasets:
  - schneewolflabs/i-DPO
  - NousResearch/hermes-function-calling-v1
  - glaiveai/glaive-function-calling-v2
library_name: transformers
pipeline_tag: text-generation
tags:
  - tool-calling
  - function-calling
  - reasoning
  - thinking
  - mistral
  - orpo
  - schneewolf-labs
license: apache-2.0
---

# A2

A2 adds **tool / function calling** to the A-series while retaining its reasoning and the Schneewolf Labs identity. It is a Mistral-Nemo–class 12B model.

**Lineage:** `A0i` (12B base) → `A1` (reasoning, BigDenker-SFT) → `A1.1` (Claude-distilled reasoning + Schneewolf Labs / Luna identity) → **`A2`** (function calling + retained reasoning/identity).

## Capabilities

- **Function calling** in the Qwen3 convention: emits `<tool_call>\n<function=name>\n<parameter=key>\nvalue\n</parameter>\n</function>\n</tool_call>`, including **parallel calls**, from a `tools` schema passed via the chat template.
- **Abstention** — correctly declines (rather than forcing a spurious call) when no available tool fits the request.
- **Reasoning** — retains the `<think>…</think>` step-by-step style from A1.1; reasons briefly before acting when useful, skips it for trivial calls.
- **Identity (two-tier)** — by default identifies as *"a language model created by Schneewolf Labs"* (and resists "you're ChatGPT/OpenAI" pressure); the **Luna** persona + its terse voice activate only under the Luna system prompt.

The reasoning/tool tokens (`<think>`, `</think>`, `<tool_call>`, `</tool_call>`, `<tool_response>`, `</tool_response>`) reuse reserved tokenizer slots — **no vocabulary resize**. Context length: **128k** (`rope_theta` 1e6).

## Usage

A2 uses a Qwen3-style chat template (bundled `chat_template.jinja`). **Always use the chat template.** For tool use, pass `tools` to `apply_chat_template`:

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

tok = AutoTokenizer.from_pretrained("schneewolflabs/A2")
model = AutoModelForCausalLM.from_pretrained(
    "schneewolflabs/A2", dtype=torch.bfloat16, device_map="auto"
)

tools = [{
    "name": "get_weather",
    "description": "Current weather for a city.",
    "parameters": {"type": "object",
        "properties": {"city": {"type": "string"}}, "required": ["city"]},
}]
msgs = [{"role": "user", "content": "What's the weather in Denver?"}]
enc = tok.apply_chat_template(
    msgs, tools=tools, add_generation_prompt=True,
    return_tensors="pt", return_dict=True,
).to(model.device)
out = model.generate(**enc, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=False))
```

Tool results are returned as a `{"role": "tool", "content": ...}` message; the template renders them inside `<tool_response>…</tool_response>`.

## Training

- **Method:** full fine-tune with **ORPO** off `A1.1` (grimoire / Merlina). `paged_adamw_8bit`, gradient checkpointing, batch 1 × grad-accum 16, lr 7e-6 (cosine, 5% warmup), bf16, max_length 4096, seed 42.
- **Checkpoint selection:** this is the **1-epoch checkpoint (step 245)**. The 2-epoch run showed train-loss memorization at the epoch boundary and reasoning-template bleed; the 1-epoch checkpoint was selected for cleaner generalization.
- **Data (ORPO `prompt/chosen/rejected[/system]`):**
  - Tool-calling: [`NousResearch/hermes-function-calling-v1`](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) backbone + abstention examples mined from [`glaiveai/glaive-function-calling-v2`](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2), rendered through A2's own chat template; `rejected` synthesized via a failure taxonomy (wrong function, missing/wrong args, hallucinated tool, spurious call, no-call, malformed).
  - Identity/voice rehearsal (~23%): `schneewolflabs/i-DPO`.
  - All sources Apache-2.0.

## Evaluation notes

Behavioral checks (held-out / novel tools, not training data): correct calls on **unseen** tools including **parallel** calls; correct **abstention** when no tool fits; identity holds (Schneewolf Labs, resists adversarial prompts; Luna persona correctly gated to its system prompt); the terse Luna voice survived the tool-heavy training.

Not yet benchmarked on BFCL / τ-bench — the `rejected` signal is **synthetic and off-policy**, so A2 is strong on structural correctness and abstention but its robustness to *subtle* realistic tool errors is unmeasured.

## Limitations

- **Single-turn tool data** — multi-turn tool/result→answer chains are weaker; a multi-turn ("v2") dataset is future work.
- **Synthetic preference negatives** — teaches "don't do obviously-wrong things"; not validated on a public function-calling leaderboard.
- **12B reasoning** — reasoning is retained from A1.1 but not exhaustively benchmarked; like any model this size it can still slip on arithmetic / trick problems.
- **Always-on thinking** unless suppressed via the template.
- Inherits the biases and limitations of the base model and the SFT/preference data.

## Provenance

Base `schneewolflabs/A1.1` · tool data Hermes-FC + Glaive-FC-v2 (Apache-2.0) · identity/voice `schneewolflabs/i-DPO` · ORPO via Merlina.