EviWeave-VisualPlanner-8B

EviWeave-VisualPlanner-8B

This submission-stage model package contains the complete PEFT/LoRA adapter used by EviWeave's structured visual planner. It does not contain the Qwen3-8B base weights or the 16 GB merged checkpoint. Obtain the base model separately from Qwen/Qwen3-8B under its original terms.

Included

• adapter_model.safetensors: complete learned LoRA parameters (rank 32).
• adapter_config.json: portable PEFT configuration with the public base-model ID.
• metadata/training_audit.json: canonical-target and repair-row audit.
• metadata/leakage_report.json: source/task/claim isolation audit.
• release_manifest.json: checksum and explicit release boundary.

Load with Transformers and PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, "/path/to/EviWeave-VisualPlanner-8B")
model.eval()

The canonical VisualPlan prompt/compiler and verifier are in the separate code package. This adapter is intended for structured visual planning, not as a stand-alone factual extractor or general report writer. Generated plans must still pass deterministic schema, grounding, fidelity, and rendering checks.

Training summary

The planner used 6,904 first-stage SFT examples over 2,040 source-disjoint tasks (5,399 final-plan and 1,505 verifier-passing repair rows), followed by a 1,920-example Boundary SFT stage (480 final-plan and 1,440 repair rows).

Limitations

The package omits training data, trainer state, merged weights, and external API models. It therefore supports inference and model inspection but not complete training reproduction during peer review.