What this model does

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

markdown
Copy code
Input:  Go program source + partial trace (GoStart, GoBlock, GoUnblock, GoCreate, GoEnd, GoSched events)
Output: {"event_type": "GoBlock", "goroutine_id": 3, "reasoning": "...", "confidence": "high"}

Training

Results (Phase 12)

Note: The 56% Gemini baseline is not a direct comparison — it used a much larger model with no fine-tuning. The relevant comparison is Qwen2.5-Coder-1.5B zero-shot vs fine-tuned, which has not yet been measured. This is planned as the next evaluation step. Investigation into the 40.2% result is ongoing — likely cause is metadata fields (concurrency_pattern, nondeterminism) missing from the JSONL.

This is an experimental research checkpoint published for reproducibility. See the GitHub repo and dataset for full context.

Usage

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "kavirubc/weave-ccwm-qwen2.5-coder-1.5b-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-1.5B-Instruct")

Or use the eval script from the repo:

bash
Copy code
uv run python scripts/run_eval.py \
    --adapter kavirubc/weave-ccwm-qwen2.5-coder-1.5b-lora \
    --val_file dataset/output/kaggle_upload/val_point_dups.jsonl

Citation

bibtex
Copy code
@misc{weave2026,
  author = {Hapuarachchi, Kaviru},
  title  = {Weave: Concurrent Code World Models},
  year   = {2026},
  url    = {https://github.com/kaviru2/Weave}
}