Replace with GDPO checkpoint (qwen36-27b-GDPO_u2f1_f1ur_h200, step 328)

README.md

@@ -1,28 +1,30 @@
 ---
 license: apache-2.0
 base_model: Qwen/Qwen3.6-27B
-tags: [debugging, code, reinforcement-learning, grpo, precisecoder]
+tags: [debugging, code, reinforcement-learning, gdpo, precisecoder]
 ---
 # PreciseCoder-27B
 
 Qwen3.6-27B post-trained with reinforcement learning for **precise debugging**: given buggy Python
 code, produce a minimal, correct fix (PreciseCoder project). Best-validation checkpoint, global step 328 (2 epochs).
 
-**Recipe.** GRPO from the base model (no SFT warm-start) with reward = (unit + 0.5·F1) / 1.5, where F1 is the
-edit-block precision/recall F1 against the ground-truth diff (PDB evaluator), on the PDB train split (buggy inputs only).
-Trained with verl on 2 nodes (n=6 rollouts, temperature 1.0, lr 1e-6, batch 32, max response 6144 tokens,
-KL loss 0.005, rollout TP=2). Minimal-edit prompt template.
+**Recipe.** GDPO from the base model (no SFT warm-start): two reward signals, unit-test pass (weight 2.0) and
+unit-gated F1 (weight 1.0), each normalized within the rollout group before being combined into the advantage.
+F1 is the edit-block precision/recall F1 against the ground-truth diff (PDB evaluator), counted only when the
+unit tests pass. Trained on the PDB train split (buggy inputs only) with verl on 1 node of 8×H200 (n=6 rollouts,
+temperature 1.0, lr 1e-6, batch 32, max response 6144 tokens, KL loss 0.005, rollout TP=2, token-level rollout
+importance sampling). Minimal-edit prompt template.
 
 **Eval** (PDB benchmark, BigCodeBench + LiveCodeBench debugging tasks):
 
 | eval | result |
 |---|---|
-| pdb_test (buggy inputs, 4k think + 2k answer) | unit 0.777 · precision 0.784 · recall 0.850 · edit lines 2.7 |
-| gt_holdout (1512 clean inputs, 16k think + 4k answer) | no-op 47.2% · damage 10.2% · unit pass 89.7% |
-| PDB-Wild (228 SWE-smith repo-level bugs) | unit 0.531 · precision 0.544 · recall 0.622 · F1 0.559 · edit lines 9.4 |
+| pdb_test (buggy inputs, 4k think + 2k answer) | unit 0.793 · precision 0.792 · recall 0.867 · F1 0.810 · edit lines 2.6 |
+| gt_holdout (1512 clean inputs, 16k think + 4k answer) | no-op 45.8% · damage 11.1% · unit pass 89.0% |
+| PDB-Wild (228 SWE-smith repo-level bugs) | unit 0.522 · precision 0.551 · recall 0.629 · F1 0.566 · edit lines 9.2 |
 
 `no-op` = share of clean (bug-free) inputs returned unchanged; `damage` = share broken by an edit.
-Computed over answers with an extractable code block.
 
 Load with `transformers` / vLLM (>=0.17) as a standard Qwen3.6 model (thinking enabled by default).
-Source: verl experiment `qwen36-27b-GRPO_0.5F1` (Delta, Sep 2026).
+Source: verl experiment `qwen36-27b-GDPO_u2f1_f1ur_h200` (Delta, Oct 2026). The previous release
+(GRPO, reward (unit + 0.5·F1)/1.5) is at revision `82cf7e4`.

model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6d3f53ea3d5156d9207e1c9bf77813a6b601f44ab87669709da82dad04aae773
-size 49905840616
+oid sha256:6587bf50bd8bd2e141d53e5730795fc448435b2b1d3d7e3304a8a01a23c6f5e9
+size 49839168128

model-00002-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1deb0689ad3e4a4b0190f0cfeba0ea5a5992f00154b95bb1a874fdadad616b49
-size 4807768096
+oid sha256:550c0545da23ee1cefc322153a5df755c921e876a692a9b4d3710033c419865e
+size 4874440624