---
base_model: mistralai/Ministral-3-3B-Instruct-2512-BF16
library_name: transformers
pipeline_tag: text-generation
tags:
- llama
- ministral
- text-only
---

# Ministral-3 3B Instruct BF16 Llama Text

Text-only Llama-compatible conversion of
`mistralai/Ministral-3-3B-Instruct-2512-BF16`.

What changed:

- dropped vision and multimodal projector tensors
- stripped the Mistral3 `language_model.` wrapper from text weights
- wrote a plain `LlamaForCausalLM` config
- kept tokenizer assets and chat template with the corrected regex
- kept chat formatting in `chat_template.jinja`
- removed the strict user/assistant alternation assertion from the template
- left tokenizer loading on the generic fast backend, not `LlamaTokenizerFast`

Verification:

- reference: `mistralai/Ministral-3-3B-Instruct-2512-BF16`
- candidate: this checkpoint
- dataset: `/home/alvion/valve/services/training/datasets/think2-2025-12-07_gpt-5.4_reasoning.jsonl`
- rows: 3
- max length: 512
- tokenizer IDs: identical
- worst KL: 0
- worst logit diff: 0
- plain candidate `AutoTokenizer` matched the fixed reference tokenizer

Conversion tools: https://github.com/cascade-tech-ai/mistral-convert