pourasrp/star_v1

STAR v1: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT

Authors: Chaitanya Medidar, Manasi Ghalsasi, Pouras Patil, Rushil Patil (Vivekanand Education Society's Institute of Technology, Mumbai)

DOI: 10.57967/hf/11003

These are the release artefacts for STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT (Medidar, Ghalsasi, Patil and Patil, 2026). Code: https://github.com/chaitanyamedidar/STAR

STAR is a selective SFT recipe that turns coding-agent recoveries into the supervision signal. It works in four steps:

1. Detect failures from execution signals in tool output, using a deterministic detector and a six-type codebook.
2. Type each failure, and check every type against a hand-labelled precision gate before any training.
3. Force-keep the recovery span after each gated failure. On the remaining candidate tokens, keep only those whose base-model per-token loss is at or above the per-trajectory median. Weights are binary.
4. Train a LoRA student on Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218) and compare it with pre-declared ablation arms.

Stage-1 gate outcome

The rule was set in advance: a type passes if precision is at least 0.80 and the Wilson 95% lower bound is at least 0.70. Only T4 (test regression) passed. T5 and T6 failed, were excluded from training, and are reported. Source: labels/stage1_gate_labels.json, records/star/gates/stage1.FAIL and stage1.T4.PASS.

type	n	YES	precision	Wilson 95%	verdict
T4	145	140	0.966	0.922–0.985	PASS
T5	90	6	0.067	0.031–0.138	FAIL
T6	142	60	0.423	0.344–0.505	FAIL

Arms

Seed 42, 500 LoRA steps each. Only A0, A5 and A2 were trained.

arm	definition (from the repo README / weights manifest)	adapter
A0	baseline: w = 1 on every candidate token; never sees the mask	pourasrp/star-A0
A5	STAR: gated T4 recovery spans force-kept, plus the median base-loss cut on the rest	pourasrp/star-A5
A2	utility-only: the same median base-loss cut, no force-keep	pourasrp/star-A2
A1	random-k, matched to A5's retention (weights only, not trained)	–
A4	ungated control: all detector-proposed T4/T5/T6 spans regardless of the gate (weights only, not trained)	–
A5-corr	an A5 variant that also force-keeps only T4 spans; see build_weights.py and weights_manifest.json for the exact rule (weights only, not trained)	–

Layout

• adapters/{A0,A5,A2}/: adapter_model.safetensors and adapter_config.json (PEFT LoRA, r=64, alpha=128)
• weights/: w_<arm>.npz binary per-token loss weights and weights_manifest.json. The manifest's w_sha256 hashes the arrays, not the npz files.
• labels/: Stage-1 hand labels (per labeller and resolved), gate labels, and annotations.jsonl (turn indices and kinds only)
• records/: gate records, Stage-1 import log, Tier-2 eval summaries and decision rules, SHA256SUMS, split id lists, scoring manifest
• scripts/: reproduction scripts, keeping their repo paths under scripts/star/...
• manifests/: per-arm training run manifests

This dataset contains no trajectory text and no SWE-Lego source code. Every input is rebuilt from upstream by instance_id.

Regenerating the inputs

Upstream data: SWE-Lego Real (SWE-Lego/SWE-Lego-Real-Data), revision 879c96b91dfa321b4d1c621d149a9200f8a1d118, resolved == True. Training ids are in records/star/labelling/split/train_pool_ids.txt. Run these from the root of the GitHub repo (commands taken from its README):

pip install -r star/training/requirements-train.txt numpy==2.2.6 pyarrow==25.0.1
python star/scoring/score_base_loss.py --full
python star/training/annotate_messages.py --out star/training/annotations/annotations.jsonl
python star/training/build_weights.py --annotations star/training/annotations/annotations.jsonl \
  --stage1-pass star/gates/stage1.T4.PASS --labels star/artifacts/results/stage1_gate_labels.json \
  --seed 42 --arms A0,A1,A2,A4,A5,A5-corr --out-dir star/training/weights
python star/training/train_lora.py --arm A5 --seed 42 --out-root <checkpoints_root> \
  --s-per-step <measured_s> --rate-inr-h <instance_rate> --balance-inr <balance>

Quick start: rebuild inputs

Run these from the root of the GitHub repo. The commands are copied from its README.

1. Install the pinned environment:

   pip install -r star/training/requirements-train.txt numpy==2.2.6 pyarrow==25.0.1
   

2. Compute the base-model per-token loss (writes star/scores/base/<id>.npz and MANIFEST.json):

   python star/scoring/score_base_loss.py --full
   

3. Build the turn/span annotations:

   python star/training/annotate_messages.py --out star/training/annotations/annotations.jsonl
   

4. Build the weight files for every arm:

   python star/training/build_weights.py --annotations star/training/annotations/annotations.jsonl \
     --stage1-pass star/gates/stage1.T4.PASS --labels star/artifacts/results/stage1_gate_labels.json \
     --seed 42 --arms A0,A1,A2,A4,A5,A5-corr --out-dir star/training/weights
   

5. (Optional) Train one arm with LoRA:

   python star/training/train_lora.py --arm A5 --seed 42 --out-root <checkpoints_root> \
     --s-per-step <measured_s> --rate-inr-h <instance_rate> --balance-inr <balance>
   

Note on the weight files

Each w_<arm>.npz holds binary per-token loss masks over 2,480 trajectories, keyed by instance_id (ids) plus token offsets (offsets, 2,481 boundaries into the flat w/fk arrays of 40,619,597 tokens). They line up only with the exact Qwen/Qwen3-8B tokenisation (revision b968826d9c46dd6066d109eabc6255188de91218) and the chat template used in the paper: tokenize_common.py sha256 a3d24851a8d584be8b044ad9ee2d1f7c2fcca7f06e65821d12934a9eacaa8bce and chat_template_sha256 a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8 (see weights/weights_manifest.json, manifests/*_run_manifest.json and records/star/scores/base/MANIFEST.json). Inputs rebuilt with any other tokeniser or template will not align.

Loading an adapter

from huggingface_hub import snapshot_download
from peft import PeftModel
from transformers import AutoModelForCausalLM
path = snapshot_download("pourasrp/star_v1", repo_type="dataset", allow_patterns="adapters/A5/*")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218")
model = PeftModel.from_pretrained(base, path, subfolder="adapters/A5")
# or load from the model repo: PeftModel.from_pretrained(base, "pourasrp/star-A5")

SHA256

file	sha256
adapters/A0/adapter_model.safetensors	dd410293b37002fa725a993f87bead22697ed3490b3fa8c7effe6b3a25c6c9a4
adapters/A5/adapter_model.safetensors	144ad8681ad879eefc0495ea305cc70b432bce26e260ee985ea2554b59590056
adapters/A2/adapter_model.safetensors	4dfe92988fc312d53e8196986e74fda7070f22eef30b8e55b77589079e782cf8
weights/w_A0.npz	ffad447aef36a064cff5dfed9f4a695b0c2171109d4a6c71d43bbcac800a560d
weights/w_A1.npz	fc6a43988f474711b0fffe719896fcc94b47075ce7dcd0127ce4388e72a74141
weights/w_A2.npz	0652bbb493334ddb76dd4e0e761918ba16833ce35ba02ba8a7781f0074aa4b1a
weights/w_A4.npz	8e8be06c281ea97f65eb81a9d2935d41673c0fb7daea0cb8741422351a1bd5b7
weights/w_A5.npz	f921f7a905d47e22ce434b6b3e2559458dc50697cfcae440d00261439dfe7466
weights/w_A5-corr.npz	0e29840c349ff2601cccd2b46609b35b15f43a175340feefbdb493085bb5875b
weights/weights_manifest.json	d36616f86e4d89388d4e0393130a8f20712abe01a7b70556937145f2d6355f86
labels/annotations.jsonl	7a8e3873b5ac0212e1e2a677d421b9f86878e7e95c074ebddd780058b9f81d6b
labels/stage1_gate_labels.json	6ff59614740e37ee3e920095bd1c99f2192c331df24024ad040c19c40146ddc9
adapters/A0/adapter_config.json	5d96cdc754568f51bea79d0fc75d22fabc5242b0cea455189fcc20824f954341
adapters/A5/adapter_config.json	ee6070f2ae37e6c3f9e89e7cbeb7c2f5b30af87e141b32735b18fcbb668313be
adapters/A2/adapter_config.json	1448204e3648947def36dc5e38f69d4b55d124a95256cba00066a8a5c27d7414
labels/stage1_labels_heldout.jsonl	11797adaab4e2a50af7aabafb856723c870fcb9d9301e9bf89132548228f21f1
labels/stage1_labels_heldout_resolved.jsonl	87c252dbf1f071ebb5adb86a23fb3b1df2152d463a940428def51ac7437fbd61

License

Apache-2.0 (see LICENSE, NOTICE). The base model Qwen/Qwen3-8B (Apache-2.0) is not included. Code inside the upstream SWE-Lego trajectories stays under each upstream repository's own license and is not redistributed here. Citation: CITATION.cff.

How to cite

DOI: 10.57967/hf/11003

@misc{star_v1_2026,
  title     = {STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT (dataset and adapters)},
  author    = {Medidar, Chaitanya and Ghalsasi, Manasi and Patil, Pouras and Patil, Rushil},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/11003},
  url       = {https://huggingface.co/datasets/pourasrp/star_v1}
}

Medidar, C., Ghalsasi, M., Patil, P., & Patil, R. (2026). STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT (dataset and adapters) [Data set]. Hugging Face. https://doi.org/10.57967/hf/11003