pourasrp/star_v1

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

Release artifacts (seed 42). Code: https://github.com/chaitanyamedidar/STAR

• A0/, A5/ (STAR), A2/: LoRA adapters for Qwen/Qwen3-8B. Load one with PeftModel.from_pretrained(base, "pourasrp/star_v1", subfolder="A5").
• weights/: per-token loss-weight files for all six arms, plus the manifest, annotations, and labels.

STAR arm A5 (seed 42) — LoRA adapter

Part of STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT. Code, paper materials and manifests: https://github.com/chaitanyamedidar/STAR

• Arm A5: STAR: T4 recovery spans force-kept + median base-loss cut on the rest
• Base model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), not included; LoRA r=64, alpha=128, all attention+MLP projections.
• Seed: 42. The training run manifest is run_manifest.json. Per-token weight files are in pourasrp/star-weights.

file	sha256
adapter_model.safetensors	144ad8681ad879eefc0495ea305cc70b432bce26e260ee985ea2554b59590056
adapter_config.json	ee6070f2ae37e6c3f9e89e7cbeb7c2f5b30af87e141b32735b18fcbb668313be

Verify with sha256sum adapter_model.safetensors.

from peft import PeftModel; from transformers import AutoModelForCausalLM
m = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B"), "pourasrp/star-A5")

License: Apache-2.0 (matches the GitHub repo). Training data came from SWE-Lego Real. No trajectory text is redistributed here.

STAR arm A0 (seed 42) — LoRA adapter

Part of STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT. Code, paper materials and manifests: https://github.com/chaitanyamedidar/STAR

• Arm A0: baseline: w = 1 on every candidate token; never sees the mask
• Base model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), not included; LoRA r=64, alpha=128, all attention+MLP projections.
• Seed: 42. The training run manifest is run_manifest.json. Per-token weight files are in pourasrp/star-weights.

file	sha256
adapter_model.safetensors	dd410293b37002fa725a993f87bead22697ed3490b3fa8c7effe6b3a25c6c9a4
adapter_config.json	5d96cdc754568f51bea79d0fc75d22fabc5242b0cea455189fcc20824f954341

Verify with sha256sum adapter_model.safetensors.

from peft import PeftModel; from transformers import AutoModelForCausalLM
m = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B"), "pourasrp/star-A0")

License: Apache-2.0 (matches the GitHub repo). Training data came from SWE-Lego Real. No trajectory text is redistributed here.

STAR arm A2 (seed 42) — LoRA adapter

Part of STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT. Code, paper materials and manifests: https://github.com/chaitanyamedidar/STAR

• Arm A2: utility-only: the same median base-loss cut, no force-keep (ablation)
• Base model: Qwen/Qwen3-8B (revision b968826d9c46dd6066d109eabc6255188de91218), not included; LoRA r=64, alpha=128, all attention+MLP projections.
• Seed: 42. The training run manifest is run_manifest.json. Per-token weight files are in pourasrp/star-weights.

file	sha256
adapter_model.safetensors	4dfe92988fc312d53e8196986e74fda7070f22eef30b8e55b77589079e782cf8
adapter_config.json	1448204e3648947def36dc5e38f69d4b55d124a95256cba00066a8a5c27d7414

Verify with sha256sum adapter_model.safetensors.

from peft import PeftModel; from transformers import AutoModelForCausalLM
m = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B"), "pourasrp/star-A2")

License: Apache-2.0 (matches the GitHub repo). Training data came from SWE-Lego Real. No trajectory text is redistributed here.

STAR per-token training weights (seed 42)

These are the artifacts for STAR: Execution-Grounded, Typed Recovery Supervision for Coding-Agent SFT. Code: https://github.com/chaitanyamedidar/STAR

• w_<arm>.npz: binary per-token loss weights for each pre-declared arm. A0 = baseline (all ones). A1 = random-k matched to A5. A2 = utility-only median base-loss cut. A4 = ungated T4/T5/T6 control. A5 = STAR (gated T4 recovery spans force-kept plus utility cut). A5-corr = the corrective-only variant. Only the A0, A5 and A2 adapters were trained: star-A0, star-A5, star-A2.
• weights_manifest.json: build manifest. Its w_sha256 values hash the weight arrays, not the npz files.
• annotations.jsonl: output of annotate_messages.py. It holds only message indices and turn kinds, with no text.
• stage1_labels_heldout*.jsonl, stage1_gate_labels.json: Stage-1 human labels and the gate result.

No trajectory text and no SWE-Lego source code are included. Weights are aligned to Qwen/Qwen3-8B tokenisation of SWE-Lego Real, rebuilt with the repo scripts.

file	sha256
LICENSE	cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30
NOTICE	b433188ea29b37eb2f608f1ca70837f0e10a7b09af86cd3f5004591e36b0821c
annotations.jsonl	7a8e3873b5ac0212e1e2a677d421b9f86878e7e95c074ebddd780058b9f81d6b
stage1_gate_labels.json	6ff59614740e37ee3e920095bd1c99f2192c331df24024ad040c19c40146ddc9
stage1_labels_heldout.jsonl	11797adaab4e2a50af7aabafb856723c870fcb9d9301e9bf89132548228f21f1
stage1_labels_heldout_resolved.jsonl	87c252dbf1f071ebb5adb86a23fb3b1df2152d463a940428def51ac7437fbd61
w_A0.npz	ffad447aef36a064cff5dfed9f4a695b0c2171109d4a6c71d43bbcac800a560d
w_A1.npz	fc6a43988f474711b0fffe719896fcc94b47075ce7dcd0127ce4388e72a74141
w_A2.npz	0652bbb493334ddb76dd4e0e761918ba16833ce35ba02ba8a7781f0074aa4b1a
w_A4.npz	8e8be06c281ea97f65eb81a9d2935d41673c0fb7daea0cb8741422351a1bd5b7
w_A5-corr.npz	0e29840c349ff2601cccd2b46609b35b15f43a175340feefbdb493085bb5875b
w_A5.npz	f921f7a905d47e22ce434b6b3e2559458dc50697cfcae440d00261439dfe7466
weights_manifest.json	d36616f86e4d89388d4e0393130a8f20712abe01a7b70556937145f2d6355f86