---
license: odc-by
language:
- en
task_categories:
- text-generation
tags:
- instruction-following
- ifeval
- ifbench
- sft
- lfm2
- lfm2.5
size_categories:
- 10K<n<100K
source_datasets:
- UniLu/IF_multi_constraints_upto5_SFT
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# IF_multi_constraints_upto5_SFT → LFM2.5 chat format

A derivative of [UniLu/IF_multi_constraints_upto5_SFT](https://huggingface.co/datasets/UniLu/IF_multi_constraints_upto5_SFT) (odc-by)
normalized for fine-tuning **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|>
```

SFT-ready precise-instruction-following pairs: the allenai IF-RLVR prompts answered by Gemma-4-31B-it and filtered with the official IFBench verifier (loose 1.0 / strict ≥ 0.8, ≤ 1024 tokens). `extra` keeps `strict_score`, `loose_score`, `response_length_words`.

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 |
| --- | --- |
| `{prompt, response}` → user/assistant messages | Chat format. |
| 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. |
| Exact-duplicate tool entries deduplicated | Template hygiene. |
| Dropped: calls to undeclared tools, unparseable arguments, argument names that are not identifiers or are Python keywords (`from`, `class`), rows over 8192 tokens, rows with no assistant tokens | Training hygiene. |

Nothing was rephrased, re-generated, or re-labelled.

## Columns

| Column | Type | Description |
| --- | --- | --- |
| `id`, `source`, `split`, `license` | str | provenance |
| `messages` | str (JSON) | canonical OpenAI-style messages; `tool_calls` arguments are dicts. **Feed this + `tools` to `apply_chat_template`** to re-render with any LFM template version. |
| `tools` | str (JSON) | OpenAI-style tool schemas |
| `text` | str | fully rendered LFM2.5 conversation, BOS included |
| `prompt_only` | bool | true when the row is a prompt ending in the generation prompt (no assistant tokens) |
| `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 `""` |
| `extra` | str (JSON) | source-specific fields: `key`, `strict_score`, `loose_score`, `response_length_words` |

## Stats

| Split | Read | Kept | Rows with renamed tools | Dropped (reason=count) |
| --- | ---: | ---: | ---: | --- |
| `train` | 42,161 | 42,161 | 0 |  |

## Training notes

- Use assistant-only loss: `apply_chat_template(messages, tools=tools, tokenize=True, return_assistant_tokens_mask=True)`.
- Prefer re-rendering from `messages`/`tools` over training on `text` if your template differs.
- Serving-side parsers must accept JSON literals (`true`, `null`, nested `{}`/`[]`) inside Pythonic calls.

## Citation

Please cite the upstream dataset: https://huggingface.co/datasets/UniLu/IF_multi_constraints_upto5_SFT