Weave-CCWM — Qwen2.5-Coder-7B Trajectory LoRA (Phase 24)

A LoRA adapter fine-tuned on Weave-Bench using trajectory-level training for next-scheduler-event prediction in concurrent Go programs. Part of the Weave project on Concurrent Code World Models (CCWM).

This model is trained on WeaveChan/WeaveMutex-instrumented programs (Phase 21+ dataset), enabling GoUnblock event prediction beyond the 0% information-theoretic floor of uninstrumented traces.

Results (Phase 24, in-distribution 545-example traj val set)

Event	Correct	Total	Accuracy
GoStart	140	211	66.4%
GoBlock	144	176	81.8%
GoCreate	11	56	19.6%
GoUnblock	9	35	25.7%
GoEnd	0	25	0%
GoSched	0	42	0%
Overall	304	545	55.8%

What this model does

Given a concurrent Go program and a partial execution trace (goroutine scheduler events), predict the next scheduler event:

Input:  Go source + partial trace (GoStart, GoBlock, GoUnblock, GoCreate, GoEnd, GoSched)
Output: {"event_type": "GoBlock", "goroutine_id": 3}

Training

Setting	Value
Base model	`Qwen/Qwen2.5-Coder-7B-Instruct`
Method	Unsloth + QLoRA, trajectory-level
Dataset	kavirubc/weave-bench (`data/train_trajectory.jsonl`)
Epochs	3
train_loss	0.02501
GPU	RTX 4000 Ada (20GB)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype=torch.float16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "kavirubc/weave-ccwm-qwen25coder-7b-traj-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

Citation

@misc{weave2026,
  author = {Hapuarachchi, Kaviru},
  title  = {Weave: Concurrent Code World Models},
  year   = {2026},
  url    = {https://arxiv.org/abs/2606.17508}
}