Reforge-1.5B β€” Python Code Repair

A trained LoRA adapter for Qwen2.5-Coder-1.5B-Instruct that proposes repairs to Python functions while preserving their names, signatures and stated requirements. This is the retained balanced-short-v1-seed42-01 checkpoint at step120. The base revision is pinned to 2e1fd397ee46e1388853d2af2c993145b0f1098a.

This repository now contains a Python repair model, not an electricity-price forecaster. The repository was renamed from chronos-2-de-lu-pregate; the old URL redirects here. The previous Chronos release remains available at revision 9ffcc6a064d27eb2c7657d3763554782710ceaca. Its forecasting results do not apply to this model.

What it does

Provide buggy Python code and a clear description of its intended behavior. The model generates a proposed corrected function. It is a repair specialist trained on authored synthetic examples; it is not established as a general coding upgrade. Outputs require review. The supplied inference helper prints code and never executes it or overwrites your input file.

Recorded results and limitations

These are saved results for this exact checkpoint, not new measurements for this release. Development panels were reused, include correlated examples, and are not independent confirmation of generalization.

Condition Base passed Adapter passed
Authored repair development, 72 tasks 34/72 65/72
Public development, 72 tasks, original three assertions only 39/72 44/72
Small public MBPP+ development screen, 24 tasks 10/24 14/24
External Original HumanEval, 164 tasks 107/164 21/164
External strict HumanEval+ (base AND plus), 164 tasks 100/164 18/164

The external assessment showed a strong regression. The narrow repair gains do not establish broader coding improvement. The HumanEval comparison used matched greedy768-token decoding and local OS sandbox caps of 60 CPU seconds, 65 wall seconds and 4096 MiB sampled RSS per group; this is not exact official CLI or leaderboard resource parity. Public pretraining overlap is possible. Finite static source reviews cannot prove correctness or noninterference. No OPAL, statistical-significance, isolated training-mechanism or deployment-readiness claim follows. The newer MBPP+ assessment03 evaluated a different, update-regularized adapter and its scores must not be attributed to this checkpoint.

The original experiment documentation rejected this adapter as a general coding replacement. It is distributed here as the explicitly selected retained repair checkpoint, with the negative results preserved. Research remains stopped. Machine-readable counts are in results.json.

Use locally

This repository ships the adapter and tokenizer, not another copy of the 1.5B base weights. The first load obtains the pinned Qwen base. No paid API or hosted inference service is included. A Hugging Face model page is not a hosted chat app.

python -m pip install -r requirements.txt

Download this repository to a local directory with hf download, or clone it with Git LFS, then run its helper:

hf download AryanSudhirDev/Reforge-1.5B --local-dir ./python-repair
python ./python-repair/inference.py \
  --code ./buggy.py \
  --requirements "Describe the required behavior and constraints here" \
  --device cpu

On an Apple Silicon Mac, use --device mps for bfloat16 inference. CPU uses float32. Add --offline once the base is cached. The helper uses greedy decoding, seed42, at most 384 new tokens and a 3072-token prompt ceiling; it rejects long inputs rather than silently truncating. This new free-form usage prompt is a convenience interface, not a newly evaluated benchmark condition.

Load with Transformers and PEFT

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

repo = "AryanSudhirDev/Reforge-1.5B"
base = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
revision = "2e1fd397ee46e1388853d2af2c993145b0f1098a"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=False)
model = AutoModelForCausalLM.from_pretrained(
    base, revision=revision, torch_dtype=torch.float32, trust_remote_code=False,
    attn_implementation="eager",
)
model = PeftModel.from_pretrained(model, repo, is_trainable=False).eval()

Use the tokenizer's user/assistant chat template to submit a repair prompt. Never execute untrusted generated code directly on your host.

Files and provenance

  • adapter_model.safetensors: byte-exact trained adapter weights.
  • adapter_config.json: portable public base-model name and pinned revision; the remaining training adapter settings are unchanged.
  • Tokenizer files: copied from the pinned Qwen base.
  • inference.py: repair CLI; generates text only, never runs generated code.
  • training_metadata.json, results.json, release_manifest.json: identity, recorded evidence and file hashes. No optimizer state, private tests, hidden inputs, canonical answers or benchmark task bodies are distributed.

LoRA rank8, alpha16, dropout0; attention q/k/v/o projections; training seed42; 120 updates. The original base model and tokenizer are by the Qwen team / Alibaba Cloud. This adapter does not claim authorship of Qwen. The upstream Apache-2.0 license is preserved in LICENSE and attribution in NOTICE.

Downloads last month
673
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for AryanSudhirDev/Reforge-1.5B

Adapter
(239)
this model