README.md

---
license: cc-by-4.0
language:
- en
task_categories:
- text-generation
tags:
- tool-use
- function-calling
- agentic
- lfm2
- lfm2.5
- liquid-ai
- pythonic-tool-calls
size_categories:
- 100K<n<1M
source_datasets:
- nvidia/Nemotron-SFT-Agentic-v2
configs:
- config_name: default
  data_files:
  - split: tool_calling
    path: data/tool_calling-*.parquet
  - split: interactive_agent
    path: data/interactive_agent-*.parquet
  - split: search
    path: data/search-*.parquet
---

# Nemotron-SFT-Agentic-v2 → LFM2.5 Pythonic tool-call format

A derivative of [nvidia/Nemotron-SFT-Agentic-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2)
(CC-BY-4.0) normalized for supervised fine-tuning of **Liquid AI LFM2 / LFM2.5** models, whose
native tool-call format is Pythonic:

```
<|im_start|>assistant
<|tool_call_start|>[get_weather(location='Paris, France', unit='celsius')]<|tool_call_end|><|im_end|>
```

Every row was rendered through the official `LiquidAI/LFM2.5-VL-3B` chat template (identical
to the LFM2.5 text models' template for text-only input) and **round-trip verified**: each rendered
Pythonic call was parsed back and compared to the source call's name and arguments. Rows that
did not round-trip exactly were dropped.

## What changed vs. the source

| Transform | Why |
| --- | --- |
| `tool_calls[].function.arguments` JSON **string → dict** | The LFM template raises on string arguments. |
| `reasoning_content` and inline `<think>…</think>` removed | LFM2.5 (non-Thinking) models answer directly. |
| Tool names sanitized to Python identifiers (`web-search` → `web_search`), consistently in `tools` and calls; originals kept in `renamed_tools` | Pythonic call syntax requires identifiers. |
| Empty system turns, legacy `function_call`, trailing non-assistant turns removed | Template hygiene. |
| Dropped: rows with `filter_reason` set, calls to undeclared tools, unparseable arguments, argument names that are Python keywords (`from`, `class` — unrepresentable as Pythonic kwargs), rows over 8192 tokens, rows with no assistant tokens | Training hygiene. |

Nothing was rephrased, re-generated, or re-labelled. Teacher-model provenance is preserved per row.

## Columns

| Column | Type | Description |
| --- | --- | --- |
| `id` | str | source uuid |
| `source`, `split`, `license` | str | provenance |
| `domain`, `teacher_model` | str | from source where present |
| `messages` | str (JSON) | canonical OpenAI-style messages; `tool_calls` arguments are dicts. **Feed this + `tools` to `apply_chat_template`** for training with any LFM template version. |
| `tools` | str (JSON) | OpenAI-style tool schemas |
| `text` | str | fully rendered LFM2.5 conversation, BOS included — ready for packing |
| `n_turns`, `n_tool_calls`, `n_tools`, `n_tokens`, `n_assistant_tokens` | int | sizes (LFM2.5 tokenizer) |
| `renamed_tools` | str (JSON) | `{original: sanitized}` when any tool was renamed, else `""` |

## Stats

| Split | Read | Kept | Rows with renamed tools | Dropped (reason=count) |
| --- | ---: | ---: | ---: | --- |
| `tool_calling` | 707,052 | 603,007 | 27,281 | bad_arg_name=1,364, call_unknown_tool=6,783, empty_assistant=5,651, keyword_arg_name=119, roundtrip_parse=12, too_long=83,909, tool_name_collision=6,207 |
| `interactive_agent` | 278,880 | 278,570 | 0 | too_long=310 |
| `search` | 5,968 | 287 | 287 | too_long=5,681 |

## Training notes

- Use assistant-only loss. The LFM template has `{% generation %}` markers, so
  `apply_chat_template(messages, tools=tools, tokenize=True, return_assistant_tokens_mask=True)` gives the mask directly.
- Prefer re-rendering from `messages`/`tools` over training on `text` if you fine-tune a model whose template differs.
- Serving-side parsers must accept JSON literals (`true`, `null`, nested `{}`/`[]`) inside Pythonic calls — the template emits them via `tojson` for non-string values.


## Citation

Please cite the upstream dataset:
NVIDIA, *Nemotron-SFT-Agentic-v2*, 2025. https://huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2