Upload folder using huggingface_hub

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/README.md

@@ -0,0 +1,5 @@
+# Issue 32 True Issue 28 Reproduction Artifacts
+
+This folder contains the rebuilt issue #28 EN->PT k=160k LoRA rank-ladder artifacts: ReLP mask, PEFT adapters, generation dumps, XCOMET summaries, logs, summaries, specs, and reusable scripts.
+
+Upstream HY-MT/XCOMET weights, Hugging Face caches, merged full checkpoints, API keys, and service tokens are intentionally excluded.

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/SHA256SUMS

@@ -0,0 +1,63 @@
+eecc85d1f56120e85e1695297132bedadc146aca0e5d2bdf025ab296c3ed3262  ./README.md
+8fb4cc48f8daabbe72742c11ea8c58d4557656a51c527aa9985c685efc6f1669  ./SHA256SUMS
+e2874b7c335668eef712ec27fe30338e7e88c32257c85ae6720903cec0449d65  ./dumps/base_k160000/fixed_k160000.jsonl
+614863ee0b3907f0549e9405637a241b1c9733abb96005c36e40ee221180d0db  ./dumps/base_k160000/no_mask.jsonl
+65b5704a1d6175bda3495336150d68f8c70110acea5ea6bff1c7ff834d981a87  ./dumps/r2_k160000/fixed_k160000.jsonl
+094af49f65afbfa854cd3e0411316ee248e3f8e7f8adb17156bb1d152e1f4d80  ./dumps/r2_k160000/no_mask.jsonl
+0257f0fdba506031b5f40cd74d02fd05f6afacedc76cbb28e7e85838fd443a9e  ./dumps/r8_k160000/fixed_k160000.jsonl
+97d4102b760f17eb8fa7e006215e260d3c465f5462a0d175e32559361fb2e1f5  ./dumps/r8_k160000/no_mask.jsonl
+d6ff1f143328336b20b2af2ed7a9ba7912db269f170f945471df09df62a79e46  ./eval/base_k160000_masks.json
+a5fbc42e83d1d379778c7d31fe5b18e4b4c69b76788d66d27a9bb1bb6e451cdb  ./eval/r2_k160000_masks.json
+be0f076d74288fffdbfab088edac678fabbbdc2ad0d7544d88ab3559bd76a3ba  ./eval/r8_k160000_masks.json
+be2218696f32bef9eeb5eced61d332e4f16d6f1043b40000243fb1c158b45c38  ./logs/download_hy_lora_conditions.log
+a6c88766189b0f6ae07f14ca2ba0325b80d7d59ae76f8f7b2d7904ad8cc21ab9  ./logs/eval_base_k160000.log
+b3067b1d32fd92aa4a7d0e43e67d08f2f6d73f468f343fa281457f9e58248afd  ./logs/eval_r2_k160000.log
+56eff871c43d71e9424375c461d35eff607b8b67c34c7ed5d21dde208b08a6f6  ./logs/eval_r8_k160000.log
+3e5eccf3effd87010bcaed7e3f0b933620dcd9d49860b50629821b205099fef2  ./logs/runner.outer.log
+a790eb859e37b46bbae6a2cb2ab9b7e265632d341d2490fa050cae8c4576e597  ./logs/xcomet_base_fixed_k160000.log
+441ab94ac405c30069e15784b41d100fe9724df85e24c46a53e86eb2c7a9c4ca  ./logs/xcomet_r2_fixed_k160000.log
+6762fc1302822d9409fb2fcfdb1282afba6dc7c42c766d27a3a92bd79d58a19e  ./logs/xcomet_r8_fixed_k160000.log
+c736705c39b7d15b553582b0abea6b71ba8f181f39a6d22d980150c55b275053  ./manifest.json
+37d6425e72bde1446eb404d0847de8d0d17857b951bc92f1b2eb8fa90224ef7a  ./notes/issue32_true28_repro.md
+d24585e5853518e2635ceefd158f678bf9af0cf322f541f86fea420edc6b27a9  ./scripts/evaluate_translation_adapter_masks.py
+9fcfd4090f86b1141571a54aa4bf7b990efd43f75879f0959d7b66661395a5ce  ./scripts/issue32_hy_lora_conditions_repro_runner.sh
+8415dbf932f91996f2bb786c7a72529d2dc30c4008982fdb7d4a6907b16c9dee  ./scripts/issue32_true28_repro_runner.sh
+e0e901f8c3f1620327cdcad37aeaf067900ddf25e752d9b7d373f000b5280d54  ./scripts/package_issue32_true28_repro_hf_upload.sh
+4b2f81f341f6f747994e9172fde8969a1df84b3c9d94d782a892cb9e3a9428ff  ./scripts/summarize_issue32_true28_repro.py
+7a085e85ff6883a887ef3c77685013bda5d93dcafb397254ada71f2799a0d150  ./source_artifacts/README.md
+2ffdd8cca915620f13b64d70960056cecafaf78bcd4bdba9b658f9418eeeef79  ./source_artifacts/low_rank_lens/k160_r2/adapter/README.md
+6f6ae26678f0f880d624c48fb79cc624cfbf6ad2a63a01fa0bdc5e67121c32d2  ./source_artifacts/low_rank_lens/k160_r2/adapter/adapter_config.json
+571d954ae125763dd73dae82fbc6c45335d15e965e8286e87cd7d6d5e089508c  ./source_artifacts/low_rank_lens/k160_r2/adapter/adapter_model.safetensors
+b7491ec0e9c869dfce20f2176758099bf248d979dd05530ede99deb21698acee  ./source_artifacts/low_rank_lens/k160_r2/adapter/chat_template.jinja
+b475bbef1b0b2fd57dcb865332b546475bd1ede2deb3bb91bafd0c047a8a530a  ./source_artifacts/low_rank_lens/k160_r2/adapter/tokenizer.json
+5abf46ce322b931371ede40634ee8c3c694202503e9f0a2eb2ec9761cbbe738e  ./source_artifacts/low_rank_lens/k160_r2/adapter/tokenizer_config.json
+5136715624a75e3964c34c266df14de3bd97c7e2fac33e5581074ac7c28fd719  ./source_artifacts/low_rank_lens/k160_r2/config.json
+480782657bf4890009dd5ffb9512e54e83fb6a3077fc973f449b22c130a0f96f  ./source_artifacts/low_rank_lens/k160_r2/train_summary.json
+2ffdd8cca915620f13b64d70960056cecafaf78bcd4bdba9b658f9418eeeef79  ./source_artifacts/low_rank_lens/k160_r8/adapter/README.md
+66315394f48a67e84ab75ce2b88a148763fd24c8c8b2ee7a64cbbe14c44783a7  ./source_artifacts/low_rank_lens/k160_r8/adapter/adapter_config.json
+409d0932242a888ed6f8703f15e5c3ed2a7e0fae3b20eb64bc88b33c27e18e7f  ./source_artifacts/low_rank_lens/k160_r8/adapter/adapter_model.safetensors
+b7491ec0e9c869dfce20f2176758099bf248d979dd05530ede99deb21698acee  ./source_artifacts/low_rank_lens/k160_r8/adapter/chat_template.jinja
+b475bbef1b0b2fd57dcb865332b546475bd1ede2deb3bb91bafd0c047a8a530a  ./source_artifacts/low_rank_lens/k160_r8/adapter/tokenizer.json
+5abf46ce322b931371ede40634ee8c3c694202503e9f0a2eb2ec9761cbbe738e  ./source_artifacts/low_rank_lens/k160_r8/adapter/tokenizer_config.json
+cf405a9daba066d2f6259126a09164ff18122c2e1099c1741f38689dc4628e6c  ./source_artifacts/low_rank_lens/k160_r8/config.json
+55be4dfeedcf429d60cc51524ddbf0435df00119b54ffffcc3f3ccbdd622bca9  ./source_artifacts/low_rank_lens/k160_r8/train_summary.json
+b537b3d0388c97ef3fa4fb7dce17aeb125c90295084141dcfd467e5f1ff29639  ./source_artifacts/manifest.json
+2ffdd8cca915620f13b64d70960056cecafaf78bcd4bdba9b658f9418eeeef79  ./source_artifacts/masked_kl/k160_r32/adapter/README.md
+a85238203fb95699c97604830d43745846f0777f8a7124efac90adc65c91534b  ./source_artifacts/masked_kl/k160_r32/adapter/adapter_config.json
+447196e8e6fb3fd89bbba8550122853a1d01adb1eeb3e145845b64cd3a7310aa  ./source_artifacts/masked_kl/k160_r32/adapter/adapter_model.safetensors
+b7491ec0e9c869dfce20f2176758099bf248d979dd05530ede99deb21698acee  ./source_artifacts/masked_kl/k160_r32/adapter/chat_template.jinja
+b475bbef1b0b2fd57dcb865332b546475bd1ede2deb3bb91bafd0c047a8a530a  ./source_artifacts/masked_kl/k160_r32/adapter/tokenizer.json
+5abf46ce322b931371ede40634ee8c3c694202503e9f0a2eb2ec9761cbbe738e  ./source_artifacts/masked_kl/k160_r32/adapter/tokenizer_config.json
+1515d92d9b83b55250fe031e083e6e32c13b03ac6c6b17accbb5bd72b44df574  ./source_artifacts/masked_kl/k160_r32/config.json
+86cdbb4d7dc526ec7cf339170eb006306e3f81775a1532af589a27f4b42671dc  ./source_artifacts/masked_kl/k160_r32/train_summary.json
+7f44da6aa7c02bb7d891b7c537de51dffa5cbf7f41ffc3e0433fa3ce25914de6  ./source_artifacts/masks/base_attr/relp_k160000.full.npz
+ed7b3aa6cc21de3b2e3f4931eba9256f450c7ba73313ae4fa90f55c9aa50d934  ./spec/issue32_hy_lora_conditions_repro.json
+49a88d8b08bf5513b9ca45ddaa381f15dc7fe9f55c7195d9c299483c4cd23246  ./spec/issue32_true28_repro_enpt.json
+b6b2eb7c5cd8cf2f37ca3106c798a1668d7ea32e75978fc0ca1777b25fe43282  ./summaries/issue32_true28_repro_summary.json
+0e9197ca775597a1cc8288086c49e5e538f0bec043676ef381caf2100bb2d0ad  ./summaries/issue32_true28_repro_summary.md
+0bd7ce7f536edb020a3c12c83285a6c07c7f18ae5a3216e0c64a7589221e35c8  ./xcomet/base_fixed_k160000.json
+e9bf4d5ebfe80c6f2516a0e39e9c218a26e5476ca40381ca9df8e2c3c13fbc62  ./xcomet/base_fixed_k160000.scored_pool.jsonl
+cb15c0d38d03d87a02d52f6392e4b9f81360bb34f60ce52d9a773bdd3b4f4aaf  ./xcomet/r2_fixed_k160000.json
+4f470707d6b421b1a608b1ed6cdde1551f9a052c6f5e563e19ca9d51215b5d3d  ./xcomet/r2_fixed_k160000.scored_pool.jsonl
+3ce7abf11c8017d4608705890c491667b2e2f1303745da39fc7deaf269d98d97  ./xcomet/r8_fixed_k160000.json
+007d0df12e8be3acbb527d3c56fcccaa507289b422aa19c5996c588e074592f3  ./xcomet/r8_fixed_k160000.scored_pool.jsonl

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/eval/base_k160000_masks.json

@@ -0,0 +1,30 @@
+{
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "adapter": null,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "n_examples": 1012,
+  "n_calib": 64,
+  "results": {
+    "no_mask": {
+      "scores": {
+        "chrFpp": 54.1532647306025,
+        "chrF": 56.86009355829216,
+        "BLEU": 25.694258665462407,
+        "n": 1012
+      },
+      "kept": -1
+    },
+    "fixed_k160000": {
+      "scores": {
+        "chrFpp": 46.967458026096146,
+        "chrF": 49.828345760177974,
+        "BLEU": 18.370739044824525,
+        "n": 1012
+      },
+      "kept": 160000,
+      "mask_path": "/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+      "elapsed_s": 56.12802314758301
+    }
+  }
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/eval/r2_k160000_masks.json

@@ -0,0 +1,30 @@
+{
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "adapter": "/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r2/adapter",
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "n_examples": 1012,
+  "n_calib": 64,
+  "results": {
+    "no_mask": {
+      "scores": {
+        "chrFpp": 54.2900767781082,
+        "chrF": 57.09539358850508,
+        "BLEU": 25.89748559065704,
+        "n": 1012
+      },
+      "kept": -1
+    },
+    "fixed_k160000": {
+      "scores": {
+        "chrFpp": 51.21753789709872,
+        "chrF": 54.0653544449977,
+        "BLEU": 22.47180926693969,
+        "n": 1012
+      },
+      "kept": 160000,
+      "mask_path": "/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+      "elapsed_s": 58.085695028305054
+    }
+  }
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/eval/r8_k160000_masks.json

@@ -0,0 +1,30 @@
+{
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "adapter": "/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r8/adapter",
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "n_examples": 1012,
+  "n_calib": 64,
+  "results": {
+    "no_mask": {
+      "scores": {
+        "chrFpp": 53.682599818043684,
+        "chrF": 56.50196836912086,
+        "BLEU": 24.96002699511978,
+        "n": 1012
+      },
+      "kept": -1
+    },
+    "fixed_k160000": {
+      "scores": {
+        "chrFpp": 51.31005902924941,
+        "chrF": 54.10140531214584,
+        "BLEU": 22.6611180101592,
+        "n": 1012
+      },
+      "kept": 160000,
+      "mask_path": "/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz",
+      "elapsed_s": 58.91963839530945
+    }
+  }
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/download_hy_lora_conditions.log

@@ -0,0 +1,4 @@
+
+Download complete. Moving file to /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+
+/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/eval_base_k160000.log

@@ -0,0 +1,12 @@
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.1532647306025, 'chrF': 56.86009355829216, 'BLEU': 25.694258665462407, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 46.967458026096146, 'chrF': 49.828345760177974, 'BLEU': 18.370739044824525, 'n': 1012} (56.1s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/base_k160000_masks.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/eval_r2_k160000.log

@@ -0,0 +1,14 @@
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] adapter /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r2/adapter
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.2900767781082, 'chrF': 57.09539358850508, 'BLEU': 25.89748559065704, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 51.21753789709872, 'chrF': 54.0653544449977, 'BLEU': 22.47180926693969, 'n': 1012} (58.1s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/r2_k160000_masks.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/eval_r8_k160000.log

@@ -0,0 +1,14 @@
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] adapter /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r8/adapter
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 53.682599818043684, 'chrF': 56.50196836912086, 'BLEU': 24.96002699511978, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 51.31005902924941, 'chrF': 54.10140531214584, 'BLEU': 22.6611180101592, 'n': 1012} (58.9s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/r8_k160000_masks.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/runner.outer.log

@@ -0,0 +1,248 @@
+
+Download complete. Moving file to /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+
+/root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.1532647306025, 'chrF': 56.86009355829216, 'BLEU': 25.694258665462407, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 46.967458026096146, 'chrF': 49.828345760177974, 'BLEU': 18.370739044824525, 'n': 1012} (56.1s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/base_k160000_masks.json
+./issue32_hy_lora_conditions_repro_runner.sh: line 190: unexpected EOF while looking for matching `"'
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] tokenizer from tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] base model tencent/HY-MT1.5-1.8B
+[load] adapter /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r2/adapter
+[load] adapter /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/low_rank_lens/k160_r8/adapter
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[merge] merging adapter in memory
+[data] eval_rows=1012
+[mean] building mean cache from 64 prompts
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+[eval] no-mask
+The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
+     no_mask: {'chrFpp': 54.2900767781082, 'chrF': 57.09539358850508, 'BLEU': 25.89748559065704, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     no_mask: {'chrFpp': 53.682599818043684, 'chrF': 56.50196836912086, 'BLEU': 24.96002699511978, 'n': 1012}
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/no_mask.jsonl
+[eval] mask=fixed_k160000 from /root/runs/issue32_hy_lora_conditions_repro/hy_lora_conditions/masks/base_attr/relp_k160000.full.npz
+     fixed_k160000: kept=160000 {'chrFpp': 51.21753789709872, 'chrF': 54.0653544449977, 'BLEU': 22.47180926693969, 'n': 1012} (58.1s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/r2_k160000_masks.json
+     fixed_k160000: kept=160000 {'chrFpp': 51.31005902924941, 'chrF': 54.10140531214584, 'BLEU': 22.6611180101592, 'n': 1012} (58.9s)
+     dumped 1012 hyps -> /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/fixed_k160000.jsonl
+[done] wrote /root/runs/issue32_hy_lora_conditions_repro/eval/r8_k160000_masks.json
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/fixed_k160000.jsonl
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/fixed_k160000.jsonl
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.8013 throughput=4.05 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.5670 throughput=3.87 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.8058 throughput=4.13 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.6006 throughput=3.91 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.8094 throughput=4.14 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.6058 throughput=3.92 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.8078 throughput=4.12 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.6068 throughput=4.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.8057 throughput=4.13 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.6155 throughput=4.01 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.8045 throughput=4.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.6166 throughput=3.97 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.8037 throughput=4.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.6255 throughput=3.96 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 1012/1012 mean=0.8031 throughput=4.06 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r2_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r2_fixed_k160000.json
+[chunk] 1012/1012 mean=0.6278 throughput=3.91 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/base_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/base_fixed_k160000.json
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/fixed_k160000.jsonl
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.8164 throughput=4.01 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.8263 throughput=4.05 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.8262 throughput=4.06 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.8286 throughput=4.04 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.8205 throughput=4.07 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.8164 throughput=4.02 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.8150 throughput=4.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 1012/1012 mean=0.8146 throughput=3.98 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r8_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r8_fixed_k160000.json
+{
+  "out_json": "/root/runs/issue32_hy_lora_conditions_repro/summaries/issue32_true28_repro_summary.json",
+  "out_md": "/root/runs/issue32_hy_lora_conditions_repro/summaries/issue32_true28_repro_summary.md",
+  "rows": 4
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/xcomet_base_fixed_k160000.log

@@ -0,0 +1,66 @@
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/base_k160000/fixed_k160000.jsonl
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.5670 throughput=3.87 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.6006 throughput=3.91 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.6058 throughput=3.92 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.6068 throughput=4.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.6155 throughput=4.01 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.6166 throughput=3.97 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.6255 throughput=3.96 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 1012/1012 mean=0.6278 throughput=3.91 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/base_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/base_fixed_k160000.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/xcomet_r2_fixed_k160000.log

@@ -0,0 +1,66 @@
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/r2_k160000/fixed_k160000.jsonl
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.8013 throughput=4.05 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.8058 throughput=4.13 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.8094 throughput=4.14 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.8078 throughput=4.12 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.8057 throughput=4.13 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.8045 throughput=4.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.8037 throughput=4.10 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [1]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 1012/1012 mean=0.8031 throughput=4.06 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r2_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r2_fixed_k160000.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/logs/xcomet_r8_fixed_k160000.log

@@ -0,0 +1,66 @@
+[load] 1012 rows from /root/runs/issue32_hy_lora_conditions_repro/dumps/r8_k160000/fixed_k160000.jsonl
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/torchmetrics/utilities/imports.py:23: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import DistributionNotFound, get_distribution
+
+[comet] loading Unbabel/XCOMET-XXL from /root/.cache/huggingface/hub/models--Unbabel--XCOMET-XXL/snapshots/873bac1b1c461e410c4a6e379f6790d3d1c7c214/checkpoints/model.ckpt
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/core/saving.py:197: Found keys that are not in the model state dict but in the checkpoint: ['encoder.model.embeddings.position_ids']
+[comet] scoring 1012 rows batch=8 chunk=128
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+You are using a CUDA device ('NVIDIA RTX PRO 6000 Blackwell Server Edition') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 128/1012 mean=0.8164 throughput=4.01 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 256/1012 mean=0.8263 throughput=4.05 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 384/1012 mean=0.8262 throughput=4.06 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 512/1012 mean=0.8286 throughput=4.04 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 640/1012 mean=0.8205 throughput=4.07 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 768/1012 mean=0.8164 throughput=4.02 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 896/1012 mean=0.8150 throughput=4.03 seg/s
+GPU available: True (cuda), used: True
+TPU available: False, using: 0 TPU cores
+💡 Tip: For seamless cloud logging and experiment tracking, try installing [litlogger](https://pypi.org/project/litlogger/) to enable LitLogger, which logs metrics and artifacts automatically to the Lightning Experiments platform.
+💡 Tip: For seamless cloud uploads and versioning, try installing [litmodels](https://pypi.org/project/litmodels/) to enable LitModelCheckpoint, which syncs automatically with the Lightning model registry.
+LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
+/root/work/circuit-shotting/.venv/lib/python3.12/site-packages/pytorch_lightning/utilities/_pytree.py:21: `isinstance(treespec, LeafSpec)` is deprecated, use `isinstance(treespec, TreeSpec) and treespec.is_leaf()` instead.
+[chunk] 1012/1012 mean=0.8146 throughput=3.98 seg/s
+[save] scored pool -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r8_fixed_k160000.scored_pool.jsonl
+[save] summary -> /root/runs/issue32_hy_lora_conditions_repro/xcomet/r8_fixed_k160000.json

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/manifest.json

@@ -0,0 +1,84 @@
+{
+  "issue": 32,
+  "task": "True issue #28 EN->PT low-rank LoRA k=160k reproduction",
+  "run_root": "/root/runs/issue32_hy_lora_conditions_repro",
+  "upload_prefix": "issue32_hy_lora_conditions_repro_20260513T141918Z",
+  "source_artifact_repo": "Occupying-Mars/hy-lora-conditions",
+  "source_commit": "11a0d84",
+  "summary": {
+    "issue": 32,
+    "run_root": "/root/runs/issue32_hy_lora_conditions_repro",
+    "k": 160000,
+    "recorded_issue28": {
+      "base_fixed_k160000": 0.6300551002,
+      "r2_fixed_k160000": 0.7655800197,
+      "r4_fixed_k160000": 0.7979727848,
+      "r8_fixed_k160000": 0.8121489377
+    },
+    "rows": [
+      {
+        "condition": "base",
+        "xcomet": 0.627828843148035,
+        "recorded_issue28_xcomet": 0.6300551002,
+        "delta_vs_recorded": -0.002226257051964997,
+        "masked_chrFpp": 46.967458026096146,
+        "masked_chrF": 49.828345760177974,
+        "masked_BLEU": 18.370739044824525,
+        "no_mask_chrFpp": 54.1532647306025
+      },
+      {
+        "condition": "r2",
+        "xcomet": 0.8031356554964314,
+        "recorded_issue28_xcomet": 0.7655800197,
+        "delta_vs_recorded": 0.0375556357964314,
+        "masked_chrFpp": 51.21753789709872,
+        "masked_chrF": 54.0653544449977,
+        "masked_BLEU": 22.47180926693969,
+        "no_mask_chrFpp": 54.2900767781082
+      },
+      {
+        "condition": "r4",
+        "xcomet": null,
+        "recorded_issue28_xcomet": 0.7979727848,
+        "delta_vs_recorded": null,
+        "masked_chrFpp": null,
+        "masked_chrF": null,
+        "masked_BLEU": null,
+        "no_mask_chrFpp": null
+      },
+      {
+        "condition": "r8",
+        "xcomet": 0.8145537328549289,
+        "recorded_issue28_xcomet": 0.8121489377,
+        "delta_vs_recorded": 0.0024047951549289737,
+        "masked_chrFpp": 51.31005902924941,
+        "masked_chrF": 54.10140531214584,
+        "masked_BLEU": 22.6611180101592,
+        "no_mask_chrFpp": 53.682599818043684
+      }
+    ],
+    "observed_ranking": [
+      "r8",
+      "r2",
+      "base"
+    ],
+    "target_ranking": [
+      "r8",
+      "r4",
+      "r2",
+      "base"
+    ],
+    "strongest_observed": {
+      "condition": "r8",
+      "xcomet": 0.8145537328549289,
+      "recorded_issue28_xcomet": 0.8121489377,
+      "delta_vs_recorded": 0.0024047951549289737,
+      "masked_chrFpp": 51.31005902924941,
+      "masked_chrF": 54.10140531214584,
+      "masked_BLEU": 22.6611180101592,
+      "no_mask_chrFpp": 53.682599818043684
+    }
+  },
+  "file_count": 60,
+  "weights_policy": "Includes PEFT adapters, but excludes upstream HY-MT/XCOMET weights, HF caches, merged full model checkpoints, API keys, and service tokens."
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/notes/issue32_true28_repro.md

@@ -0,0 +1,65 @@
+# Issue 32 True Issue 28 Reproduction Runbook
+
+This run reconstructs the strongest recorded issue #28 EN->PT result from the
+issue #23/#24/#28 science trail because the exact #28 adapter and mask bundle was
+not initially available.
+
+Update: the preserved adapter/mask bundle is now known to live at
+`Occupying-Mars/hy-lora-conditions`. That repo should be treated as the source
+of truth for reproducing #28 before any regenerated approximation. Its manifest
+contains `low_rank_lens/k160_r2`, `low_rank_lens/k160_r8`, and
+`masks/base_attr/relp_k160000.full.npz`; it does not list a
+`low_rank_lens/k160_r4` adapter.
+
+The reproduction target is the issue #28 XCOMET table:
+
+| condition | k | XCOMET |
+|---|---:|---:|
+| base fixed mask | 160000 | `0.6300551002` |
+| r2 LoRA fixed mask | 160000 | `0.7655800197` |
+| r4 LoRA fixed mask | 160000 | `0.7979727848` |
+| r8 LoRA fixed mask | 160000 | `0.8121489377` |
+
+The preserved-artifact replay uses:
+
+- source repo: `Occupying-Mars/hy-lora-conditions`
+- base model: `tencent/HY-MT1.5-1.8B`
+- held-out eval: NTREX EN->PT, 1,012 rows
+- mask: preserved `masks/base_attr/relp_k160000.full.npz`
+- available low-rank lens adapters: `low_rank_lens/k160_r2/adapter`,
+  `low_rank_lens/k160_r8/adapter`
+- replay runner: `issue32_hy_lora_conditions_repro_runner.sh`
+
+The regenerated fallback setup follows the prior issue trail:
+
+- base model: `tencent/HY-MT1.5-1.8B`
+- train/calibration source: FLORES EN->PT devtest, 1,012 rows
+- train targets: unmasked HY-MT teacher hypotheses in `model_hyp`
+- attribution source: FLORES EN->PT dev inside `attribute_translation.py`
+- held-out eval: NTREX EN->PT, 1,012 rows
+- mask: ReLP `relp_k160000.full.npz`, `n_attr=200`, `first_token_logit`
+- lesion: mean ablation at MLP `down_proj` input, selected channels live
+- rescue: all-linear rsLoRA, ranks `2,4,8`, `alpha=2*r`
+- objective: masked KL to unmasked teacher logits + CE on teacher hypotheses +
+  small unmasked KL guardrail
+
+The runner intentionally avoids saving merged full model checkpoints. It keeps
+PEFT adapters and evaluates them by merging into memory, so backups do not
+duplicate upstream HY-MT weights.
+
+Primary files:
+
+- `configs/issue32_true28_repro_enpt.json`
+- `evaluate_translation_adapter_masks.py`
+- `issue32_true28_repro_runner.sh`
+- `summarize_issue32_true28_repro.py`
+- `scripts/package_issue32_true28_repro_hf_upload.sh`
+
+Interpretation:
+
+- A strong reproduction should recover the issue #28 ordering
+  `r8 > r4 > r2 > base` and land near the recorded r8 XCOMET frontier.
+- If the reconstructed base/mask anchor is far away, the run should be treated
+  as a fresh #28-style reproduction attempt, not a byte-identical replay.
+- Only after this table is reproduced should the region-student contract be
+  compared against the strongest r8 anchor.

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/scripts/evaluate_translation_adapter_masks.py

@@ -0,0 +1,301 @@
+#!/usr/bin/env python3
+"""Evaluate HY-MT EN->PT translation masks with an optional PEFT adapter.
+
+This is the issue #28 reproduction evaluator. It mirrors
+evaluate_translation_masks.py, but loads a base model plus a saved LoRA adapter
+and merges the adapter in memory before generation. That preserves the original
+LoRA rescue inference contract without writing full merged model checkpoints to
+disk.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import numpy as np
+import sacrebleu
+import torch
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from src.circuit_tracing.ablation import MeanCache
+from translation_io import (
+    DEFAULT_PROMPT_STYLE,
+    DEFAULT_TARGET_LANGUAGE,
+    Pair,
+    apply_chat,
+    generate_translations,
+    load_flores_devtest_any,
+)
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--base-model", required=True)
+    p.add_argument("--adapter", default=None,
+                   help="Optional PEFT adapter dir. If set, it is merged in memory.")
+    p.add_argument("--mask", action="append", default=[], metavar="NAME:PATH")
+    p.add_argument("--out", required=True)
+    p.add_argument("--device", default="cuda")
+    p.add_argument("--dtype", default="bfloat16",
+                   choices=["float32", "float16", "bfloat16"])
+    p.add_argument("--target-language", default=DEFAULT_TARGET_LANGUAGE)
+    p.add_argument("--prompt-style", default=DEFAULT_PROMPT_STYLE,
+                   choices=["hy_mt", "sarvam"])
+    p.add_argument("--src-lang", default="eng_Latn")
+    p.add_argument("--tgt-lang", default="por_Latn")
+    p.add_argument("--input-jsonl", default=None)
+    p.add_argument("--n-calib", type=int, default=64)
+    p.add_argument("--max-examples", type=int, default=None)
+    p.add_argument("--batch-size", type=int, default=8)
+    p.add_argument("--max-new-tokens", type=int, default=384)
+    p.add_argument("--include-no-mask", action="store_true")
+    p.add_argument("--dump-hyps-dir", default=None)
+    p.add_argument("--dump-category", default="flores_devtest")
+    p.add_argument("--dump-tag", default="heldout")
+    return p.parse_args()
+
+
+def text_config(model):
+    return getattr(model.config, "text_config", model.config)
+
+
+def load_mask_npz(path: Path, n_layers: int, d_ffn: int) -> Dict[int, torch.Tensor]:
+    arr = np.load(path)
+    out: Dict[int, torch.Tensor] = {}
+    for layer in range(n_layers):
+        key = f"layer_{layer}"
+        if key in arr:
+            out[layer] = torch.from_numpy(arr[key]).bool()
+        else:
+            out[layer] = torch.zeros(d_ffn, dtype=torch.bool)
+    return out
+
+
+def parse_mask_specs(values: List[str]) -> List[Tuple[str, Path]]:
+    specs = []
+    for value in values:
+        if ":" not in value:
+            raise ValueError(f"--mask must be NAME:PATH, got {value!r}")
+        name, raw = value.split(":", 1)
+        specs.append((name, Path(raw)))
+    return specs
+
+
+def decoder_root(model):
+    cur = model
+    for _ in range(8):
+        if hasattr(cur, "layers"):
+            return cur
+        for attr in ("model", "language_model", "base_model"):
+            nxt = getattr(cur, attr, None)
+            if nxt is not None and nxt is not cur:
+                cur = nxt
+                break
+        else:
+            break
+    raise AttributeError("could not locate decoder .layers")
+
+
+def install_mask_hooks(model, keep_per_layer, base_per_layer):
+    hooks = []
+    root = decoder_root(model)
+    for layer_idx, layer in enumerate(root.layers):
+        keep = keep_per_layer[layer_idx]
+        base = base_per_layer[layer_idx]
+
+        def make(keep, base):
+            def hook_fn(module, args):
+                act = args[0]
+                modified = act * keep + base * (1.0 - keep)
+                return (modified,) + args[1:]
+            return hook_fn
+
+        hooks.append(layer.mlp.down_proj.register_forward_pre_hook(make(keep, base)))
+    return hooks
+
+
+def evaluate_with_mask(model, tokenizer, sources, refs, *,
+                       n_layers, d_ffn, mask, mean_cache,
+                       device, dtype, batch_size, max_new_tokens,
+                       target_language, prompt_style=DEFAULT_PROMPT_STYLE):
+    keep_per_layer = {}
+    for layer_idx in range(n_layers):
+        keep = torch.zeros(d_ffn, device=device, dtype=dtype)
+        if mask is not None:
+            mask_layer = mask[layer_idx].to(device)
+            keep[mask_layer] = 1.0
+        keep_per_layer[layer_idx] = keep.view(1, 1, -1)
+
+    base_per_layer = {}
+    for layer_idx in range(n_layers):
+        mu = mean_cache.means.get(layer_idx)
+        if mu is None:
+            base_per_layer[layer_idx] = torch.zeros(1, 1, d_ffn, device=device, dtype=dtype)
+        else:
+            base_per_layer[layer_idx] = mu.to(device=device, dtype=dtype).view(1, 1, -1)
+
+    hooks = install_mask_hooks(model, keep_per_layer, base_per_layer) if mask is not None else []
+    try:
+        hyps = generate_translations(
+            model, tokenizer, sources,
+            target_language=target_language,
+            prompt_style=prompt_style,
+            batch_size=batch_size,
+            max_new_tokens=max_new_tokens,
+            do_sample=False,
+            device=device,
+        )
+    finally:
+        for hook in hooks:
+            hook.remove()
+
+    chrfpp = sacrebleu.corpus_chrf(hyps, [refs], word_order=2).score
+    chrf = sacrebleu.corpus_chrf(hyps, [refs], word_order=0).score
+    bleu = sacrebleu.corpus_bleu(hyps, [refs]).score
+    return {"chrFpp": chrfpp, "chrF": chrf, "BLEU": bleu, "n": len(hyps)}, hyps
+
+
+def load_pairs(args: argparse.Namespace) -> list[Pair]:
+    if args.input_jsonl:
+        pairs = []
+        with open(args.input_jsonl) as fh:
+            for line in fh:
+                if not line.strip():
+                    continue
+                row = json.loads(line)
+                pairs.append(Pair(src=row["en"], tgt=row.get("pt", "")))
+                if args.max_examples and len(pairs) >= args.max_examples:
+                    break
+        return pairs
+    return load_flores_devtest_any(
+        src_lang=args.src_lang,
+        tgt_lang=args.tgt_lang,
+        max_examples=args.max_examples,
+    )
+
+
+def main() -> None:
+    args = parse_args()
+    out_path = Path(args.out)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    dtype = {"float32": torch.float32, "float16": torch.float16,
+             "bfloat16": torch.bfloat16}[args.dtype]
+
+    print(f"[load] tokenizer from {args.base_model}", flush=True)
+    tokenizer = AutoTokenizer.from_pretrained(args.base_model)
+    print(f"[load] base model {args.base_model}", flush=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        args.base_model, dtype=dtype, attn_implementation="eager",
+    ).to(args.device).eval()
+    if args.adapter:
+        print(f"[load] adapter {args.adapter}", flush=True)
+        model = PeftModel.from_pretrained(model, args.adapter)
+        print("[merge] merging adapter in memory", flush=True)
+        model = model.merge_and_unload().to(args.device).eval()
+
+    cfg = text_config(model)
+    n_layers = int(cfg.num_hidden_layers)
+    d_ffn = int(cfg.intermediate_size)
+    pairs = load_pairs(args)
+    sources = [p.src for p in pairs]
+    refs = [p.tgt for p in pairs]
+    print(f"[data] eval_rows={len(pairs)}", flush=True)
+
+    print(f"[mean] building mean cache from {args.n_calib} prompts", flush=True)
+    calib_ids = []
+    for src in sources[: args.n_calib]:
+        prompt = apply_chat(
+            tokenizer, src,
+            target_language=args.target_language,
+            add_generation_prompt=True,
+            prompt_style=args.prompt_style,
+        )
+        ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids
+        calib_ids.append(ids)
+    mean_cache = MeanCache.build(model, calib_ids, device=args.device)
+    mean_cache = MeanCache(means={k: v.to(dtype=dtype) for k, v in mean_cache.means.items()})
+
+    dump_dir = Path(args.dump_hyps_dir) if args.dump_hyps_dir else None
+    if dump_dir is not None:
+        dump_dir.mkdir(parents=True, exist_ok=True)
+
+    def dump_block(block_name: str, hyps: List[str]) -> None:
+        if dump_dir is None:
+            return
+        path = dump_dir / f"{block_name}.jsonl"
+        with path.open("w") as fh:
+            for idx, (src, ref, hyp) in enumerate(zip(sources, refs, hyps)):
+                fh.write(json.dumps({
+                    "id": idx,
+                    "en": src,
+                    "pt": ref,
+                    "model_hyp": hyp,
+                    "category": args.dump_category,
+                    "tag": args.dump_tag,
+                    "mask_name": block_name,
+                    "adapter": args.adapter,
+                }, ensure_ascii=False) + "\n")
+        print(f"     dumped {len(hyps)} hyps -> {path}", flush=True)
+
+    results = {}
+    if args.include_no_mask:
+        print("[eval] no-mask", flush=True)
+        scores, hyps = evaluate_with_mask(
+            model, tokenizer, sources, refs,
+            n_layers=n_layers, d_ffn=d_ffn,
+            mask=None, mean_cache=mean_cache,
+            device=args.device, dtype=dtype,
+            batch_size=args.batch_size,
+            max_new_tokens=args.max_new_tokens,
+            target_language=args.target_language,
+            prompt_style=args.prompt_style,
+        )
+        print(f"     no_mask: {scores}", flush=True)
+        results["no_mask"] = {"scores": scores, "kept": -1}
+        dump_block("no_mask", hyps)
+
+    for name, mask_path in parse_mask_specs(args.mask):
+        print(f"[eval] mask={name} from {mask_path}", flush=True)
+        mask = load_mask_npz(mask_path, n_layers, d_ffn)
+        kept = sum(int(m.sum().item()) for m in mask.values())
+        t0 = time.time()
+        scores, hyps = evaluate_with_mask(
+            model, tokenizer, sources, refs,
+            n_layers=n_layers, d_ffn=d_ffn,
+            mask=mask, mean_cache=mean_cache,
+            device=args.device, dtype=dtype,
+            batch_size=args.batch_size,
+            max_new_tokens=args.max_new_tokens,
+            target_language=args.target_language,
+            prompt_style=args.prompt_style,
+        )
+        elapsed = time.time() - t0
+        print(f"     {name}: kept={kept} {scores} ({elapsed:.1f}s)", flush=True)
+        results[name] = {
+            "scores": scores,
+            "kept": kept,
+            "mask_path": str(mask_path),
+            "elapsed_s": elapsed,
+        }
+        dump_block(name, hyps)
+
+    with out_path.open("w") as f:
+        json.dump({
+            "base_model": args.base_model,
+            "adapter": args.adapter,
+            "n_layers": n_layers,
+            "d_ffn": d_ffn,
+            "n_examples": len(pairs),
+            "n_calib": args.n_calib,
+            "results": results,
+        }, f, indent=2, ensure_ascii=False)
+    print(f"[done] wrote {out_path}", flush=True)
+
+
+if __name__ == "__main__":
+    main()

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/scripts/issue32_hy_lora_conditions_repro_runner.sh

@@ -0,0 +1,200 @@
+#!/usr/bin/env bash
+# Replay preserved issue #28 EN->PT LoRA-lens artifacts from HF.
+set -euo pipefail
+
+REPO_DIR="${REPO_DIR:-/root/work/circuit-shotting}"
+cd "$REPO_DIR"
+source .venv/bin/activate
+
+MODEL="${MODEL:-tencent/HY-MT1.5-1.8B}"
+ARTIFACT_REPO="${ARTIFACT_REPO:-Occupying-Mars/hy-lora-conditions}"
+RUN_ROOT="${RUN_ROOT:-/root/runs/issue32_hy_lora_conditions_repro}"
+ARTIFACT_DIR="${ARTIFACT_DIR:-$RUN_ROOT/hy_lora_conditions}"
+NTREX_JSONL="${NTREX_JSONL:-/root/runs/ntrex_eval/ntrex_en2pt.jsonl}"
+K="${K:-160000}"
+SHORT_K="${SHORT_K:-$((K / 1000))}"
+RANKS="${RANKS:-2 8}"
+GPU_LIST="${GPU_LIST:-0}"
+SCORE_AFTER="${SCORE_AFTER:-1}"
+UPLOAD_AFTER="${UPLOAD_AFTER:-0}"
+
+mkdir -p "$RUN_ROOT"/{logs,eval,dumps,xcomet,summaries} "$ARTIFACT_DIR"
+
+pick_gpu() {
+  local idx="$1"
+  python - "$GPU_LIST" "$idx" <<'PY'
+import sys
+gpus=[g for g in sys.argv[1].replace(",", " ").split() if g]
+if not gpus:
+    gpus=["0"]
+print(gpus[int(sys.argv[2]) % len(gpus)])
+PY
+}
+
+gpu_count() {
+  python - "$GPU_LIST" <<'PY'
+import sys
+gpus=[g for g in sys.argv[1].replace(",", " ").split() if g]
+print(max(len(gpus), 1))
+PY
+}
+
+wait_one() {
+  local -n _pids="$1"
+  if ((${#_pids[@]} == 0)); then
+    return 0
+  fi
+  wait "${_pids[0]}"
+  _pids=("${_pids[@]:1}")
+}
+
+if [[ ! -f "$ARTIFACT_DIR/manifest.json" ]]; then
+  uv run python - "$ARTIFACT_REPO" "$ARTIFACT_DIR" "$K" "$SHORT_K" <<'PY' \
+    2>&1 | tee "$RUN_ROOT/logs/download_hy_lora_conditions.log"
+import sys
+
+from huggingface_hub import snapshot_download
+
+repo, local_dir, k, short_k = sys.argv[1:5]
+patterns = [
+    "README.md",
+    "manifest.json",
+    f"low_rank_lens/k{short_k}_r*/adapter/*",
+    f"low_rank_lens/k{short_k}_r*/config.json",
+    f"low_rank_lens/k{short_k}_r*/train_summary.json",
+    f"masked_kl/k{short_k}_r32/adapter/*",
+    f"masked_kl/k{short_k}_r32/config.json",
+    f"masked_kl/k{short_k}_r32/train_summary.json",
+    f"masks/base_attr/relp_k{k}.full.npz",
+]
+print(snapshot_download(
+    repo_id=repo,
+    repo_type="model",
+    local_dir=local_dir,
+    allow_patterns=patterns,
+))
+PY
+fi
+
+MASK="$ARTIFACT_DIR/masks/base_attr/relp_k${K}.full.npz"
+if [[ ! -f "$MASK" ]]; then
+  echo "missing preserved mask: $MASK" >&2
+  exit 2
+fi
+
+if [[ ! -f "$NTREX_JSONL" ]]; then
+  uv run python build_ntrex_en2pt_jsonl.py 2>&1 | tee "$RUN_ROOT/logs/build_ntrex.log"
+fi
+
+if [[ ! -f "$RUN_ROOT/eval/base_k${K}_masks.json" ]]; then
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" uv run python evaluate_translation_adapter_masks.py \
+    --base-model "$MODEL" \
+    --input-jsonl "$NTREX_JSONL" \
+    --mask "fixed_k${K}:$MASK" \
+    --out "$RUN_ROOT/eval/base_k${K}_masks.json" \
+    --target-language Portuguese \
+    --include-no-mask \
+    --dump-hyps-dir "$RUN_ROOT/dumps/base_k${K}" \
+    --dump-category ntrex_test \
+    --dump-tag heldout \
+    --batch-size "${EVAL_BATCH_SIZE:-16}" \
+    --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/eval_base_k${K}.log"
+fi
+
+run_rank_eval() {
+  local r="$1"
+  local gpu="$2"
+  local adapter="$ARTIFACT_DIR/low_rank_lens/k${SHORT_K}_r${r}/adapter"
+  if [[ ! -f "$adapter/adapter_model.safetensors" ]]; then
+    echo "missing preserved rank $r adapter: $adapter" >&2
+    return 3
+  fi
+  if [[ ! -f "$RUN_ROOT/eval/r${r}_k${K}_masks.json" ]]; then
+    CUDA_VISIBLE_DEVICES="$gpu" uv run python evaluate_translation_adapter_masks.py \
+      --base-model "$MODEL" \
+      --adapter "$adapter" \
+      --input-jsonl "$NTREX_JSONL" \
+      --mask "fixed_k${K}:$MASK" \
+      --out "$RUN_ROOT/eval/r${r}_k${K}_masks.json" \
+      --target-language Portuguese \
+      --include-no-mask \
+      --dump-hyps-dir "$RUN_ROOT/dumps/r${r}_k${K}" \
+      --dump-category ntrex_test \
+      --dump-tag heldout \
+      --batch-size "${EVAL_BATCH_SIZE:-16}" \
+      --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+      2>&1 | tee "$RUN_ROOT/logs/eval_r${r}_k${K}.log"
+  fi
+}
+
+rank_idx=0
+rank_pids=()
+parallel_jobs="$(gpu_count)"
+for r in $RANKS; do
+  gpu="$(pick_gpu "$rank_idx")"
+  rank_idx=$((rank_idx + 1))
+  run_rank_eval "$r" "$gpu" &
+  rank_pids+=("$!")
+  if ((${#rank_pids[@]} >= parallel_jobs)); then
+    wait_one rank_pids
+  fi
+done
+while ((${#rank_pids[@]})); do
+  wait_one rank_pids
+done
+
+if [[ "$SCORE_AFTER" == "1" || "$SCORE_AFTER" == "true" ]]; then
+  score_specs=("base_k${K}/fixed_k${K}:base_fixed_k${K}")
+  for r in $RANKS; do
+    score_specs+=("r${r}_k${K}/fixed_k${K}:r${r}_fixed_k${K}")
+  done
+
+  run_score() {
+    local spec="$1"
+    local gpu="$2"
+    local rel="${spec%%:*}"
+    local name="${spec##*:}"
+    local in_jsonl="$RUN_ROOT/dumps/$rel.jsonl"
+    local out_json="$RUN_ROOT/xcomet/$name.json"
+    local out_pool="$RUN_ROOT/xcomet/$name.scored_pool.jsonl"
+    if [[ -f "$out_json" ]]; then
+      return 0
+    fi
+    CUDA_VISIBLE_DEVICES="$gpu" uv run python score_xcomet_pool.py \
+      --hyps-jsonl "$in_jsonl" \
+      --out-jsonl "$out_pool" \
+      --summary-json "$out_json" \
+      --comet-model "${XCOMET_MODEL:-Unbabel/XCOMET-XXL}" \
+      --system-name "$name" \
+      --batch-size "${XCOMET_BATCH_SIZE:-8}" \
+      --chunk-size "${XCOMET_CHUNK_SIZE:-128}" \
+      2>&1 | tee "$RUN_ROOT/logs/xcomet_${name}.log"
+  }
+
+  score_idx=0
+  score_pids=()
+  parallel_jobs="$(gpu_count)"
+  for spec in "${score_specs[@]}"; do
+    gpu="$(pick_gpu "$score_idx")"
+    score_idx=$((score_idx + 1))
+    run_score "$spec" "$gpu" &
+    score_pids+=("$!")
+    if ((${#score_pids[@]} >= parallel_jobs)); then
+      wait_one score_pids
+    fi
+  done
+  while ((${#score_pids[@]})); do
+    wait_one score_pids
+  done
+fi
+
+uv run python summarize_issue32_true28_repro.py \
+  --run-root "$RUN_ROOT" \
+  --k "$K" \
+  --out-json "$RUN_ROOT/summaries/issue32_true28_repro_summary.json" \
+  --out-md "$RUN_ROOT/summaries/issue32_true28_repro_summary.md"
+
+if [[ "$UPLOAD_AFTER" == "1" || "$UPLOAD_AFTER" == "true" ]]; then
+  scripts/package_issue32_true28_repro_hf_upload.sh
+fi

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/scripts/issue32_true28_repro_runner.sh

@@ -0,0 +1,220 @@
+#!/usr/bin/env bash
+# Reconstruct and rerun the strongest issue #28 EN->PT k=160k rank ladder.
+set -euo pipefail
+
+REPO_DIR="${REPO_DIR:-/root/work/circuit-shotting}"
+cd "$REPO_DIR"
+source .venv/bin/activate
+
+MODEL="${MODEL:-tencent/HY-MT1.5-1.8B}"
+RUN_ROOT="${RUN_ROOT:-/root/runs/issue32_true28_repro}"
+FLORES_JSONL="${FLORES_JSONL:-/root/runs/flores_eval/flores_en2pt_devtest.jsonl}"
+NTREX_JSONL="${NTREX_JSONL:-/root/runs/ntrex_eval/ntrex_en2pt.jsonl}"
+TRAIN_JSONL="${TRAIN_JSONL:-$RUN_ROOT/teacher_train_hyps.jsonl}"
+MASK_ROOT="${MASK_ROOT:-$RUN_ROOT/base_attr}"
+K="${K:-160000}"
+RANKS="${RANKS:-2 4 8}"
+GPU_LIST="${GPU_LIST:-0}"
+SCORE_AFTER="${SCORE_AFTER:-1}"
+UPLOAD_AFTER="${UPLOAD_AFTER:-0}"
+
+mkdir -p "$RUN_ROOT"/{logs,eval,dumps,xcomet,summaries} "$MASK_ROOT"
+
+pick_gpu() {
+  local idx="$1"
+  python - "$GPU_LIST" "$idx" <<'PY'
+import sys
+gpus=[g for g in sys.argv[1].replace(",", " ").split() if g]
+if not gpus:
+    gpus=["0"]
+print(gpus[int(sys.argv[2]) % len(gpus)])
+PY
+}
+
+gpu_count() {
+  python - "$GPU_LIST" <<'PY'
+import sys
+gpus=[g for g in sys.argv[1].replace(",", " ").split() if g]
+print(max(len(gpus), 1))
+PY
+}
+
+wait_one() {
+  local -n _pids="$1"
+  if ((${#_pids[@]} == 0)); then
+    return 0
+  fi
+  wait "${_pids[0]}"
+  _pids=("${_pids[@]:1}")
+}
+
+if [[ ! -f "$FLORES_JSONL" ]]; then
+  uv run python build_flores_en2pt_jsonl.py 2>&1 | tee "$RUN_ROOT/logs/build_flores.log"
+fi
+
+if [[ ! -f "$NTREX_JSONL" ]]; then
+  uv run python build_ntrex_en2pt_jsonl.py 2>&1 | tee "$RUN_ROOT/logs/build_ntrex.log"
+fi
+
+if [[ ! -f "$TRAIN_JSONL" ]]; then
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" uv run python gen_translations_only.py \
+    --jsonl "$FLORES_JSONL" \
+    --model "$MODEL" \
+    --out "$TRAIN_JSONL" \
+    --target-language Portuguese \
+    --batch-size "${TEACHER_BATCH_SIZE:-16}" \
+    --max-new-tokens "${TEACHER_MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/generate_teacher_hyps.log"
+fi
+
+if [[ ! -f "$MASK_ROOT/relp_k${K}.full.npz" ]]; then
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" uv run python attribute_translation.py \
+    --model "$MODEL" \
+    --out-dir "$MASK_ROOT" \
+    --target-language Portuguese \
+    --src-lang eng_Latn \
+    --tgt-lang por_Latn \
+    --n-attr "${N_ATTR:-200}" \
+    --ks "$K" \
+    --metric first_token_logit \
+    2>&1 | tee "$RUN_ROOT/logs/attribute_base_k${K}.log"
+fi
+
+MASK="$MASK_ROOT/relp_k${K}.full.npz"
+
+if [[ ! -f "$RUN_ROOT/eval/base_k${K}_masks.json" ]]; then
+  CUDA_VISIBLE_DEVICES="$(pick_gpu 0)" uv run python evaluate_translation_adapter_masks.py \
+    --base-model "$MODEL" \
+    --input-jsonl "$NTREX_JSONL" \
+    --mask "fixed_k${K}:$MASK" \
+    --out "$RUN_ROOT/eval/base_k${K}_masks.json" \
+    --target-language Portuguese \
+    --include-no-mask \
+    --dump-hyps-dir "$RUN_ROOT/dumps/base_k${K}" \
+    --dump-category ntrex_test \
+    --dump-tag heldout \
+    --batch-size "${EVAL_BATCH_SIZE:-16}" \
+    --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+    2>&1 | tee "$RUN_ROOT/logs/eval_base_k${K}.log"
+fi
+
+run_rank() {
+  local r="$1"
+  local gpu="$2"
+  alpha=$((2 * r))
+  out="$RUN_ROOT/k${K}/r${r}"
+
+  if [[ ! -f "$out/train_summary.json" ]]; then
+    mkdir -p "$out"
+    CUDA_VISIBLE_DEVICES="$gpu" uv run python train_masked_kl_conditioning.py \
+      --model "$MODEL" \
+      --jsonl "$TRAIN_JSONL" \
+      --fixed-mask "$MASK" \
+      --out-dir "$out" \
+      --target-language Portuguese \
+      --target-field model_hyp \
+      --lora-r "$r" \
+      --lora-alpha "$alpha" \
+      --batch-size "${BATCH_SIZE:-2}" \
+      --grad-accum "${GRAD_ACCUM:-8}" \
+      --epochs "${EPOCHS:-1.0}" \
+      --lr "${LR:-2e-4}" \
+      --masked-kl-beta "${MASKED_KL_BETA:-1.0}" \
+      --ce-beta "${CE_BETA:-0.2}" \
+      --unmasked-kl-beta "${UNMASKED_KL_BETA:-0.05}" \
+      --n-calib "${N_CALIB:-128}" \
+      --eval-every "${EVAL_EVERY:-10}" \
+      --no-save-merged \
+      2>&1 | tee "$RUN_ROOT/logs/train_r${r}_k${K}.log"
+  fi
+
+  if [[ ! -f "$RUN_ROOT/eval/r${r}_k${K}_masks.json" ]]; then
+    CUDA_VISIBLE_DEVICES="$gpu" uv run python evaluate_translation_adapter_masks.py \
+      --base-model "$MODEL" \
+      --adapter "$out/adapter" \
+      --input-jsonl "$NTREX_JSONL" \
+      --mask "fixed_k${K}:$MASK" \
+      --out "$RUN_ROOT/eval/r${r}_k${K}_masks.json" \
+      --target-language Portuguese \
+      --include-no-mask \
+      --dump-hyps-dir "$RUN_ROOT/dumps/r${r}_k${K}" \
+      --dump-category ntrex_test \
+      --dump-tag heldout \
+      --batch-size "${EVAL_BATCH_SIZE:-16}" \
+      --max-new-tokens "${MAX_NEW_TOKENS:-384}" \
+      2>&1 | tee "$RUN_ROOT/logs/eval_r${r}_k${K}.log"
+  fi
+}
+
+rank_index=0
+rank_pids=()
+parallel_jobs="$(gpu_count)"
+for r in $RANKS; do
+  gpu="$(pick_gpu "$rank_index")"
+  rank_index=$((rank_index + 1))
+  run_rank "$r" "$gpu" &
+  rank_pids+=("$!")
+  if ((${#rank_pids[@]} >= parallel_jobs)); then
+    wait_one rank_pids
+  fi
+done
+while ((${#rank_pids[@]})); do
+  wait_one rank_pids
+done
+
+if [[ "$SCORE_AFTER" == "1" || "$SCORE_AFTER" == "true" ]]; then
+  score_specs=("base_k${K}/fixed_k${K}:base_fixed_k${K}")
+  for r in $RANKS; do
+    score_specs+=("r${r}_k${K}/fixed_k${K}:r${r}_fixed_k${K}")
+  done
+
+  run_score() {
+    local spec="$1"
+    local gpu="$2"
+    rel="${spec%%:*}"
+    name="${spec##*:}"
+    in_jsonl="$RUN_ROOT/dumps/$rel.jsonl"
+    out_json="$RUN_ROOT/xcomet/$name.json"
+    out_pool="$RUN_ROOT/xcomet/$name.scored_pool.jsonl"
+    if [[ -f "$out_json" ]]; then
+      continue
+    fi
+    gpu="$(pick_gpu "$score_idx")"
+    score_idx=$((score_idx + 1))
+    CUDA_VISIBLE_DEVICES="$gpu" uv run python score_xcomet_pool.py \
+      --hyps-jsonl "$in_jsonl" \
+      --out-jsonl "$out_pool" \
+      --summary-json "$out_json" \
+      --comet-model "${XCOMET_MODEL:-Unbabel/XCOMET-XXL}" \
+      --system-name "$name" \
+      --batch-size "${XCOMET_BATCH_SIZE:-8}" \
+      --chunk-size "${XCOMET_CHUNK_SIZE:-128}" \
+      2>&1 | tee "$RUN_ROOT/logs/xcomet_${name}.log"
+  }
+
+  score_idx=0
+  score_pids=()
+  parallel_jobs="$(gpu_count)"
+  for spec in "${score_specs[@]}"; do
+    gpu="$(pick_gpu "$score_idx")"
+    score_idx=$((score_idx + 1))
+    run_score "$spec" "$gpu" &
+    score_pids+=("$!")
+    if ((${#score_pids[@]} >= parallel_jobs)); then
+      wait_one score_pids
+    fi
+  done
+  while ((${#score_pids[@]})); do
+    wait_one score_pids
+  done
+fi
+
+uv run python summarize_issue32_true28_repro.py \
+  --run-root "$RUN_ROOT" \
+  --k "$K" \
+  --out-json "$RUN_ROOT/summaries/issue32_true28_repro_summary.json" \
+  --out-md "$RUN_ROOT/summaries/issue32_true28_repro_summary.md"
+
+if [[ "$UPLOAD_AFTER" == "1" || "$UPLOAD_AFTER" == "true" ]]; then
+  scripts/package_issue32_true28_repro_hf_upload.sh
+fi

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/scripts/package_issue32_true28_repro_hf_upload.sh

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+# Package and upload issue #32 true issue #28 reproduction artifacts.
+set -euo pipefail
+
+RUN_ROOT="${RUN_ROOT:-/root/runs/issue32_true28_repro}"
+REPO_DIR="${REPO_DIR:-/root/work/circuit-shotting}"
+HF_REPO="${HF_REPO:-TokenBender/synth-data-en-pt-circuit}"
+UPLOAD_ROOT="${UPLOAD_ROOT:-$RUN_ROOT/hf_upload}"
+STAMP="${STAMP:-$(date -u +%Y%m%dT%H%M%SZ)}"
+UPLOAD_PREFIX="${UPLOAD_PREFIX:-issue32_true28_repro_$STAMP}"
+UPLOAD_DIR="$UPLOAD_ROOT/$UPLOAD_PREFIX"
+
+mkdir -p "$UPLOAD_DIR"/{spec,notes,scripts,base_attr,data,eval,dumps,xcomet,summaries,logs,adapters,manifests,source_artifacts}
+
+cp "$REPO_DIR/configs/issue32_true28_repro_enpt.json" "$UPLOAD_DIR/spec/"
+cp "$REPO_DIR/configs/issue32_hy_lora_conditions_repro.json" "$UPLOAD_DIR/spec/" 2>/dev/null || true
+cp "$REPO_DIR/notes/issue32_true28_repro.md" "$UPLOAD_DIR/notes/"
+cp "$REPO_DIR/evaluate_translation_adapter_masks.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/issue32_true28_repro_runner.sh" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/issue32_hy_lora_conditions_repro_runner.sh" "$UPLOAD_DIR/scripts/" 2>/dev/null || true
+cp "$REPO_DIR/summarize_issue32_true28_repro.py" "$UPLOAD_DIR/scripts/"
+cp "$REPO_DIR/scripts/package_issue32_true28_repro_hf_upload.sh" "$UPLOAD_DIR/scripts/"
+
+cp "$RUN_ROOT"/teacher_train_hyps.jsonl "$UPLOAD_DIR/data/" 2>/dev/null || true
+cp "$RUN_ROOT"/base_attr/*.npz "$UPLOAD_DIR/base_attr/" 2>/dev/null || true
+cp "$RUN_ROOT"/base_attr/*.json "$UPLOAD_DIR/base_attr/" 2>/dev/null || true
+cp "$RUN_ROOT"/eval/*.json "$UPLOAD_DIR/eval/" 2>/dev/null || true
+if [[ -d "$RUN_ROOT/dumps" ]]; then
+  (cd "$RUN_ROOT/dumps" && find . -type f -name '*.jsonl' \
+    | while read -r path; do
+        mkdir -p "$UPLOAD_DIR/dumps/$(dirname "$path")"
+        cp "$path" "$UPLOAD_DIR/dumps/$path"
+      done)
+fi
+cp "$RUN_ROOT"/xcomet/*.json "$UPLOAD_DIR/xcomet/" 2>/dev/null || true
+cp "$RUN_ROOT"/xcomet/*.jsonl "$UPLOAD_DIR/xcomet/" 2>/dev/null || true
+cp "$RUN_ROOT"/summaries/* "$UPLOAD_DIR/summaries/" 2>/dev/null || true
+cp "$RUN_ROOT"/logs/*.log "$UPLOAD_DIR/logs/" 2>/dev/null || true
+
+if [[ -d "$RUN_ROOT/k160000" ]]; then
+  find "$RUN_ROOT/k160000" -type f \( -path '*/adapter/*' -o -name 'train_summary.json' -o -name 'config.json' \) \
+    | while read -r path; do
+        rel="${path#$RUN_ROOT/}"
+        if [[ "$rel" == */merged/* ]]; then
+          continue
+        fi
+        mkdir -p "$UPLOAD_DIR/$(dirname "$rel")"
+        cp "$path" "$UPLOAD_DIR/$rel"
+      done
+fi
+
+if [[ -d "$RUN_ROOT/hy_lora_conditions" ]]; then
+  (cd "$RUN_ROOT/hy_lora_conditions" && find . -type f \
+    | while read -r path; do
+        case "$path" in
+          */.cache/*) continue ;;
+        esac
+        mkdir -p "$UPLOAD_DIR/source_artifacts/$(dirname "$path")"
+        cp "$path" "$UPLOAD_DIR/source_artifacts/$path"
+      done)
+fi
+
+python - "$UPLOAD_DIR" "$RUN_ROOT" "$UPLOAD_PREFIX" <<'PY'
+import json
+import os
+import sys
+from pathlib import Path
+
+upload = Path(sys.argv[1])
+run_root = Path(sys.argv[2])
+prefix = sys.argv[3]
+
+def maybe_json(path: Path):
+    if path.exists():
+        return json.loads(path.read_text())
+    return None
+
+manifest = {
+    "issue": 32,
+    "task": "True issue #28 EN->PT low-rank LoRA k=160k reproduction",
+    "run_root": str(run_root),
+    "upload_prefix": prefix,
+    "source_artifact_repo": os.environ.get("SOURCE_ARTIFACT_REPO") or os.environ.get("ARTIFACT_REPO"),
+    "source_commit": os.environ.get("SOURCE_COMMIT"),
+    "summary": maybe_json(run_root / "summaries" / "issue32_true28_repro_summary.json"),
+    "file_count": sum(len(files) for _, _, files in os.walk(upload)),
+    "weights_policy": "Includes PEFT adapters, but excludes upstream HY-MT/XCOMET weights, HF caches, merged full model checkpoints, API keys, and service tokens.",
+}
+(upload / "manifest.json").write_text(json.dumps(manifest, indent=2, ensure_ascii=False) + "\n")
+(upload / "README.md").write_text(
+    "# Issue 32 True Issue 28 Reproduction Artifacts\n\n"
+    "This folder contains the rebuilt issue #28 EN->PT k=160k LoRA rank-ladder "
+    "artifacts: ReLP mask, PEFT adapters, generation dumps, XCOMET summaries, "
+    "logs, summaries, specs, and reusable scripts.\n\n"
+    "Upstream HY-MT/XCOMET weights, Hugging Face caches, merged full checkpoints, "
+    "API keys, and service tokens are intentionally excluded.\n"
+)
+print(json.dumps({"upload_dir": str(upload), "prefix": prefix, "files": manifest["file_count"]}, indent=2))
+PY
+
+(cd "$UPLOAD_DIR" && find . -type f -print0 | sort -z | xargs -0 sha256sum > SHA256SUMS)
+uv run hf upload "$HF_REPO" "$UPLOAD_DIR" "$UPLOAD_PREFIX" --repo-type dataset
+echo "$UPLOAD_PREFIX"

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/scripts/summarize_issue32_true28_repro.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+"""Summarize the true issue #28 reproduction attempt for issue #32."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+
+
+RECORDED = {
+    "base_fixed_k160000": 0.6300551002,
+    "r2_fixed_k160000": 0.7655800197,
+    "r4_fixed_k160000": 0.7979727848,
+    "r8_fixed_k160000": 0.8121489377,
+}
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--run-root", type=Path, required=True)
+    p.add_argument("--k", type=int, default=160000)
+    p.add_argument("--out-json", type=Path, required=True)
+    p.add_argument("--out-md", type=Path, required=True)
+    return p.parse_args()
+
+
+def load_json(path: Path) -> dict[str, Any] | None:
+    if not path.exists():
+        return None
+    return json.loads(path.read_text())
+
+
+def metric(payload: dict[str, Any] | None, block: str, key: str) -> float | None:
+    if not payload:
+        return None
+    row = (payload.get("results") or {}).get(block) or {}
+    scores = row.get("scores") or {}
+    val = scores.get(key)
+    return float(val) if isinstance(val, (int, float)) else None
+
+
+def xcomet(path: Path) -> float | None:
+    payload = load_json(path)
+    if not payload:
+        return None
+    value = payload.get("system_score") or payload.get("system_xcomet_xxl")
+    if isinstance(value, (int, float)):
+        return float(value)
+    summary = payload.get("summary")
+    if isinstance(summary, dict):
+        value = summary.get("system_score")
+        if isinstance(value, (int, float)):
+            return float(value)
+    return None
+
+
+def main() -> None:
+    args = parse_args()
+    root = args.run_root
+    rows = []
+    labels = ["base", "r2", "r4", "r8"]
+    for label in labels:
+        if label == "base":
+            eval_path = root / "eval" / f"base_k{args.k}_masks.json"
+            xname = f"base_fixed_k{args.k}"
+        else:
+            rank = label.removeprefix("r")
+            eval_path = root / "eval" / f"r{rank}_k{args.k}_masks.json"
+            xname = f"{label}_fixed_k{args.k}"
+        payload = load_json(eval_path)
+        score = xcomet(root / "xcomet" / f"{xname}.json")
+        recorded = RECORDED.get(xname)
+        rows.append({
+            "condition": label,
+            "xcomet": score,
+            "recorded_issue28_xcomet": recorded,
+            "delta_vs_recorded": None if score is None or recorded is None else score - recorded,
+            "masked_chrFpp": metric(payload, f"fixed_k{args.k}", "chrFpp"),
+            "masked_chrF": metric(payload, f"fixed_k{args.k}", "chrF"),
+            "masked_BLEU": metric(payload, f"fixed_k{args.k}", "BLEU"),
+            "no_mask_chrFpp": metric(payload, "no_mask", "chrFpp"),
+        })
+
+    observed = [r for r in rows if r["xcomet"] is not None]
+    ranking = sorted(observed, key=lambda r: r["xcomet"], reverse=True)
+    summary = {
+        "issue": 32,
+        "run_root": str(root),
+        "k": args.k,
+        "recorded_issue28": RECORDED,
+        "rows": rows,
+        "observed_ranking": [r["condition"] for r in ranking],
+        "target_ranking": ["r8", "r4", "r2", "base"],
+        "strongest_observed": ranking[0] if ranking else None,
+    }
+
+    args.out_json.parent.mkdir(parents=True, exist_ok=True)
+    args.out_json.write_text(json.dumps(summary, indent=2, ensure_ascii=False) + "\n")
+
+    lines = [
+        "# Issue 32 True Issue 28 Reproduction Summary",
+        "",
+        f"- Run root: `{root}`",
+        f"- Budget: `k={args.k}`",
+        "- Target ranking: `r8 > r4 > r2 > base`",
+        "",
+        "| condition | XCOMET | issue #28 recorded | delta | masked chrF++ | no-mask chrF++ |",
+        "|---|---:|---:|---:|---:|---:|",
+    ]
+    for row in rows:
+        def fmt(value: float | None) -> str:
+            return "" if value is None else f"{value:.6f}"
+        lines.append(
+            "| `{}` | {} | {} | {} | {} | {} |".format(
+                row["condition"],
+                fmt(row["xcomet"]),
+                fmt(row["recorded_issue28_xcomet"]),
+                fmt(row["delta_vs_recorded"]),
+                fmt(row["masked_chrFpp"]),
+                fmt(row["no_mask_chrFpp"]),
+            )
+        )
+    if ranking:
+        lines.extend([
+            "",
+            f"Observed ranking: `{' > '.join(r['condition'] for r in ranking)}`.",
+            f"Strongest observed condition: `{ranking[0]['condition']}`.",
+        ])
+    args.out_md.parent.mkdir(parents=True, exist_ok=True)
+    args.out_md.write_text("\n".join(lines) + "\n")
+    print(json.dumps({"out_json": str(args.out_json), "out_md": str(args.out_md), "rows": len(rows)}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/README.md

@@ -0,0 +1,29 @@
+# hy-lora-conditions
+
+private en->pt adapter and mask bundle.
+
+code lives in github: https://github.com/Occupying-Mars/circuit-shotting
+
+base model: `tencent/HY-MT1.5-1.8B`
+
+task: english -> portuguese
+
+contents:
+- adapters
+- base masks
+- lens attribution masks
+- config and train summary json files
+- manifest with adapter -> mask mapping
+
+no eval dumps, no eval tables, no logs, no code snapshots, no merged full models.
+
+## artifact map
+
+| folder | rank | mask |
+|---|---:|---|
+| `low_rank_lens/k160_r2` | 2 | `masks/base_attr/relp_k160000.full.npz` |
+| `low_rank_lens/k160_r8` | 8 | `masks/base_attr/relp_k160000.full.npz` |
+| `masked_kl/k160_r32` | 32 | `masks/base_attr/relp_k160000.full.npz` |
+| `masked_kl/k120_r32` | 32 | `masks/base_attr/relp_k120000.full.npz` |
+
+see `manifest.json` for exact adapter/config/mask paths.

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/adapter/README.md

@@ -0,0 +1,207 @@
+---
+base_model: tencent/HY-MT1.5-1.8B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:tencent/HY-MT1.5-1.8B
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.19.1

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/adapter/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "tencent/HY-MT1.5-1.8B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 2,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "o_proj",
+    "up_proj",
+    "k_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": true
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:571d954ae125763dd73dae82fbc6c45335d15e965e8286e87cd7d6d5e089508c
+size 9757496

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/adapter/chat_template.jinja

@@ -0,0 +1 @@
+{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}<｜hy_begin▁of▁sentence｜>{{ system_message }}<｜hy_place▁holder▁no▁3｜>{% else %}{% set loop_messages = messages %}<｜hy_begin▁of▁sentence｜>{% endif %}{% for message in loop_messages %}{% if message['role'] == 'user' %}<｜hy_User｜>{{ message['content'] }}{% elif message['role'] == 'assistant' %}<｜hy_Assistant｜>{{ message['content'] }}<｜hy_place▁holder▁no▁2｜>{% endif %}{% endfor %}{% if add_generation_prompt %}<｜hy_Assistant｜>{% else %}<｜hy_place▁holder▁no▁8｜>{% endif %}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/adapter/tokenizer_config.json

@@ -0,0 +1,10 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<｜hy_begin▁of▁sentence｜>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<｜hy_place▁holder▁no▁2｜>",
+  "is_local": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<｜hy_▁pad▁｜>",
+  "tokenizer_class": "TokenizersBackend"
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/config.json

@@ -0,0 +1,47 @@
+{
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/repro_issues23_28/teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/repro_issues23_28/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/repro_issues23_28/issue28_k160_r2",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 42,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 8,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "lora_r": 2,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "target_modules": "all-linear",
+    "use_rslora": true,
+    "init_adapter": null,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "unmasked_kl_beta": 0.0,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 50,
+    "num_workers": 0,
+    "save_merged": true
+  },
+  "n_rows": 1012,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "total_steps": 64,
+  "warmup_steps": 3,
+  "logs": []
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r2/train_summary.json

@@ -0,0 +1,77 @@
+{
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/repro_issues23_28/teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/repro_issues23_28/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/repro_issues23_28/issue28_k160_r2",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 42,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 8,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "lora_r": 2,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "target_modules": "all-linear",
+    "use_rslora": true,
+    "init_adapter": null,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "unmasked_kl_beta": 0.0,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 50,
+    "num_workers": 0,
+    "save_merged": true
+  },
+  "n_rows": 1012,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "total_steps": 64,
+  "warmup_steps": 3,
+  "logs": [
+    {
+      "step": 1,
+      "loss": 0.8935546875,
+      "masked_kl": 0.729736328125,
+      "ce": 0.81787109375,
+      "unmasked_kl": 0.0,
+      "lr": 6.666666666666667e-05,
+      "elapsed_s": 2.5306222438812256
+    },
+    {
+      "step": 50,
+      "loss": 0.45174984056122447,
+      "masked_kl": 0.35723876953125,
+      "ce": 0.4722800741390306,
+      "unmasked_kl": 0.0,
+      "lr": 2.4886806912948035e-05,
+      "elapsed_s": 85.60768342018127
+    },
+    {
+      "step": 64,
+      "loss": 0.35251290457589285,
+      "masked_kl": 0.27527073451450895,
+      "ce": 0.38635689871651785,
+      "unmasked_kl": 0.0,
+      "lr": 0.0,
+      "elapsed_s": 108.8494644165039
+    }
+  ],
+  "elapsed_s": 108.85567808151245,
+  "merged_dir": "/root/runs/repro_issues23_28/issue28_k160_r2/merged"
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/adapter/README.md

@@ -0,0 +1,207 @@
+---
+base_model: tencent/HY-MT1.5-1.8B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:tencent/HY-MT1.5-1.8B
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.19.1

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/adapter/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "tencent/HY-MT1.5-1.8B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj",
+    "down_proj",
+    "gate_proj",
+    "k_proj",
+    "up_proj",
+    "o_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": true
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:409d0932242a888ed6f8703f15e5c3ed2a7e0fae3b20eb64bc88b33c27e18e7f
+size 38856224

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/adapter/chat_template.jinja

@@ -0,0 +1 @@
+{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}<｜hy_begin▁of▁sentence｜>{{ system_message }}<｜hy_place▁holder▁no▁3｜>{% else %}{% set loop_messages = messages %}<｜hy_begin▁of▁sentence｜>{% endif %}{% for message in loop_messages %}{% if message['role'] == 'user' %}<｜hy_User｜>{{ message['content'] }}{% elif message['role'] == 'assistant' %}<｜hy_Assistant｜>{{ message['content'] }}<｜hy_place▁holder▁no▁2｜>{% endif %}{% endfor %}{% if add_generation_prompt %}<｜hy_Assistant｜>{% else %}<｜hy_place▁holder▁no▁8｜>{% endif %}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/adapter/tokenizer_config.json

@@ -0,0 +1,10 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<｜hy_begin▁of▁sentence｜>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<｜hy_place▁holder▁no▁2｜>",
+  "is_local": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<｜hy_▁pad▁｜>",
+  "tokenizer_class": "TokenizersBackend"
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/config.json

@@ -0,0 +1,47 @@
+{
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/repro_issues23_28/teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/repro_issues23_28/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/repro_issues23_28/issue28_k160_r8",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 42,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 8,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "lora_r": 8,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "target_modules": "all-linear",
+    "use_rslora": true,
+    "init_adapter": null,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "unmasked_kl_beta": 0.0,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 50,
+    "num_workers": 0,
+    "save_merged": true
+  },
+  "n_rows": 1012,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "total_steps": 64,
+  "warmup_steps": 3,
+  "logs": []
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/low_rank_lens/k160_r8/train_summary.json

@@ -0,0 +1,77 @@
+{
+  "args": {
+    "model": "tencent/HY-MT1.5-1.8B",
+    "jsonl": "/root/runs/repro_issues23_28/teacher_train_hyps.jsonl",
+    "fixed_mask": "/root/runs/repro_issues23_28/base_attr/relp_k160000.full.npz",
+    "out_dir": "/root/runs/repro_issues23_28/issue28_k160_r8",
+    "target_field": "model_hyp",
+    "target_language": "Portuguese",
+    "prompt_style": "hy_mt",
+    "device": "cuda",
+    "dtype": "bfloat16",
+    "seed": 42,
+    "max_rows": null,
+    "max_seq_length": 1024,
+    "n_calib": 128,
+    "mean_on": "full",
+    "epochs": 1.0,
+    "max_steps": null,
+    "batch_size": 2,
+    "grad_accum": 8,
+    "lr": 0.0002,
+    "weight_decay": 0.0,
+    "warmup_ratio": 0.05,
+    "max_grad_norm": 1.0,
+    "lora_r": 8,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "target_modules": "all-linear",
+    "use_rslora": true,
+    "init_adapter": null,
+    "masked_kl_beta": 1.0,
+    "ce_beta": 0.2,
+    "unmasked_kl_beta": 0.0,
+    "kl_temperature": 1.0,
+    "kl_on": "answer",
+    "eval_every": 50,
+    "num_workers": 0,
+    "save_merged": true
+  },
+  "n_rows": 1012,
+  "n_layers": 32,
+  "d_ffn": 6144,
+  "mask_kept": 160000,
+  "total_steps": 64,
+  "warmup_steps": 3,
+  "logs": [
+    {
+      "step": 1,
+      "loss": 0.89892578125,
+      "masked_kl": 0.729736328125,
+      "ce": 0.85009765625,
+      "unmasked_kl": 0.0,
+      "lr": 6.666666666666667e-05,
+      "elapsed_s": 2.4842658042907715
+    },
+    {
+      "step": 50,
+      "loss": 0.4577499701052296,
+      "masked_kl": 0.36263477558992346,
+      "ce": 0.4758761658960459,
+      "unmasked_kl": 0.0,
+      "lr": 2.4886806912948035e-05,
+      "elapsed_s": 83.07147860527039
+    },
+    {
+      "step": 64,
+      "loss": 0.35915701729910715,
+      "masked_kl": 0.2801361083984375,
+      "ce": 0.39583042689732145,
+      "unmasked_kl": 0.0,
+      "lr": 0.0,
+      "elapsed_s": 105.1619131565094
+    }
+  ],
+  "elapsed_s": 105.1679368019104,
+  "merged_dir": "/root/runs/repro_issues23_28/issue28_k160_r8/merged"
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/manifest.json

@@ -0,0 +1,44 @@
+{
+  "created_utc": "2026-05-13T13:17:56.780574+00:00",
+  "base_model": "tencent/HY-MT1.5-1.8B",
+  "task": "en_to_pt",
+  "repo_code": "https://github.com/Occupying-Mars/circuit-shotting",
+  "artifacts": [
+    {
+      "folder": "low_rank_lens/k160_r2",
+      "adapter": "low_rank_lens/k160_r2/adapter",
+      "config": "low_rank_lens/k160_r2/config.json",
+      "train_summary": "low_rank_lens/k160_r2/train_summary.json",
+      "mask": "masks/base_attr/relp_k160000.full.npz",
+      "mask_k": 160000,
+      "lora_rank": 2
+    },
+    {
+      "folder": "low_rank_lens/k160_r8",
+      "adapter": "low_rank_lens/k160_r8/adapter",
+      "config": "low_rank_lens/k160_r8/config.json",
+      "train_summary": "low_rank_lens/k160_r8/train_summary.json",
+      "mask": "masks/base_attr/relp_k160000.full.npz",
+      "mask_k": 160000,
+      "lora_rank": 8
+    },
+    {
+      "folder": "masked_kl/k160_r32",
+      "adapter": "masked_kl/k160_r32/adapter",
+      "config": "masked_kl/k160_r32/config.json",
+      "train_summary": "masked_kl/k160_r32/train_summary.json",
+      "mask": "masks/base_attr/relp_k160000.full.npz",
+      "mask_k": 160000,
+      "lora_rank": 32
+    },
+    {
+      "folder": "masked_kl/k120_r32",
+      "adapter": "masked_kl/k120_r32/adapter",
+      "config": "masked_kl/k120_r32/config.json",
+      "train_summary": "masked_kl/k120_r32/train_summary.json",
+      "mask": "masks/base_attr/relp_k120000.full.npz",
+      "mask_k": 120000,
+      "lora_rank": 32
+    }
+  ]
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/masked_kl/k160_r32/adapter/README.md

@@ -0,0 +1,207 @@
+---
+base_model: tencent/HY-MT1.5-1.8B
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:tencent/HY-MT1.5-1.8B
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.19.1

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/masked_kl/k160_r32/adapter/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "tencent/HY-MT1.5-1.8B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "up_proj",
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "o_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": true
+}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/masked_kl/k160_r32/adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:447196e8e6fb3fd89bbba8550122853a1d01adb1eeb3e145845b64cd3a7310aa
+size 155249120

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/masked_kl/k160_r32/adapter/chat_template.jinja

@@ -0,0 +1 @@
+{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}<｜hy_begin▁of▁sentence｜>{{ system_message }}<｜hy_place▁holder▁no▁3｜>{% else %}{% set loop_messages = messages %}<｜hy_begin▁of▁sentence｜>{% endif %}{% for message in loop_messages %}{% if message['role'] == 'user' %}<｜hy_User｜>{{ message['content'] }}{% elif message['role'] == 'assistant' %}<｜hy_Assistant｜>{{ message['content'] }}<｜hy_place▁holder▁no▁2｜>{% endif %}{% endfor %}{% if add_generation_prompt %}<｜hy_Assistant｜>{% else %}<｜hy_place▁holder▁no▁8｜>{% endif %}

circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z/source_artifacts/masked_kl/k160_r32/adapter/tokenizer_config.json

@@ -0,0 +1,10 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<｜hy_begin▁of▁sentence｜>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<｜hy_place▁holder▁no▁2｜>",
+  "is_local": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<｜hy_▁pad▁｜>",
+  "tokenizer_class": "TokenizersBackend"
+}