an-sub-acc-2026/weave-ccwm-qwen2.5-coder-7b-traj-lora

Weave-CCWM — Qwen2.5-Coder-7B LoRA (Phase 13)

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

What this model does

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

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

Training

Setting	Value
Base model	Qwen/Qwen2.5-Coder-7B-Instruct
Method	Unsloth + QLoRA
Dataset	an-sub-acc-2026/weave-bench

Results (Phase 13)

Model	Accuracy	Notes
Qwen2.5-Coder-7B fine-tuned (this model)	36.2%	Phase 13 on GoKer held-out set
Qwen2.5-Coder-1.5B fine-tuned	40.2%	Phase 12 (in-distribution evaluation)

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.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "an-sub-acc-2026/weave-ccwm-qwen2.5-coder-7b-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

Or use the eval script from the repo:

uv run python scripts/run_eval_unsloth.py \
    --adapter an-sub-acc-2026/weave-ccwm-qwen2.5-coder-7b-lora \
    --val_file dataset/output/kaggle_upload/val_point_dups.jsonl

Citation

If you use this model, dataset, or codebase, please cite:

@misc{anon-artifacts,
  author       = {Anonymous},
  title        = {Weave: Dataset, Trained Adapters, and Tooling},
  howpublished = {\url{https://github.com/an-sub-acc-2026/weave-submission} and \url{https://huggingface.co/datasets/an-sub-acc-2026/weave-bench}},
  year         = {2026}
}