qwen-3.5-babel-9b-sft

qwen-3.5-babel-9b-sft

This is the 9B supervised-fine-tuned checkpoint from Babel-Formal, a research project on proof-term-to-tactic-script translation between Lean and Rocq.

The checkpoint is derived from Qwen/Qwen3.5-9B and fine-tuned on the Babel Lean/Rocq term-to-script training corpus. It is released as a text-only Qwen3_5ForCausalLM checkpoint: the original Qwen3.5 wrapper's language-model parameters and tokenizer are retained, while no vision encoder weights are included.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "theostos/qwen-3.5-babel-9b-sft"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

Use the prompting and verification pipeline in the Babel-Formal repository for Lean/Rocq translation experiments. Generated proof scripts must be checked by the target prover; the model does not guarantee correctness.

Limitations

• This is a research checkpoint specialized for formal-proof translation.
• It is not a general-purpose mathematical correctness oracle.
• Outputs can be syntactically invalid or fail verification.
• The model card does not claim that pretraining-data contamination has been excluded.

License

Apache 2.0, following the upstream Qwen3.5 checkpoint.