fern v3b: +1 epoch from v3; 26-way 0.85, xnli 0.84, multi 0.78

README.md

@@ -74,23 +74,23 @@ codes (A–Z / 0–9 / yes,no) at the answer position. See the repo for the prom
 
 Agreement = argmax matches the teacher's argmax on held-out label files.
 
-| | v3 |
+| | v3b |
 |---|---|
 | Held-out training distribution (7.7k) | 0.94 |
 | Thinking teacher, MMLU-Pro (2k, 10-way) | 0.49 |
 | MMLU-Pro accuracy (2k) | 0.42 |
-| 11–26-way choice (920) | 0.84 |
+| 11–26-way choice (920) | 0.85 |
 | XNLI, 10 languages (740) | 0.84 |
-| Synthetic requests, 6 non-English languages (1.5k) | 0.77 |
-| Synthetic requests, English (1.9k, in-distribution) | 0.89 |
+| Synthetic requests, 6 non-English languages (1.5k) | 0.78 |
+| Synthetic requests, English (1.9k, in-distribution) | 0.90 |
 | Latency, 1 / 10 questions (p50, RTX PRO 6000, bf16) | 28 / 48 ms |
 
 ## Training
 
 - Teacher: DeepSeek-V4-Flash UD-IQ2_M via llama-server, prefill only (`n_probs`), thinking
   budget 512 tokens for knowledge sources, none elsewhere.
-- Student: full fine-tune, bf16, AdamW lr 1e-5 cosine, batch 64, 2 epochs, max_len 1,024,
-  loss `KL(p_teacher ‖ p_student)` over the option codes. 379k rows (after weighting), 17.6 h.
+- Student: full fine-tune, bf16, AdamW lr 1e-5 cosine, batch 64, 2 epochs + 1 continued epoch at lr 5e-6, max_len 1,024,
+  loss `KL(p_teacher ‖ p_student)` over the option codes. 379k rows (after weighting), 17.6 h + 8.8 h.
 - Sources: HellaSwag, MNLI, XNLI, Yelp reviews, CommonsenseQA, BoolQ, ARC, TweetEval, MMLU
   (+ auxiliary_train), OpenBookQA, SciQ, GSM8K (multiple-choice form), 11–26-way variants
   padded with same-source distractors, ~40k synthetic Jev-style requests in 7 languages

eval.json

@@ -1,52 +1,52 @@
 {
-  "model": "runs/v3",
+  "model": "runs/v3b",
   "quant": null,
   "agreement": {
     "rows": 5125,
-    "agree": 0.6829268292682927,
-    "kl": 0.6334491947849932,
-    "brier": 0.2812951607993448,
+    "agree": 0.6874146341463415,
+    "kl": 0.6401720955939848,
+    "brier": 0.28186191172584546,
     "per_file": {
       "data/eval_mmlu_pro_think512.jsonl": {
         "rows": 1999,
-        "agree": 0.48924462231115556,
-        "kl": 1.3330464420592731,
-        "brier": 0.5794463507974281
+        "agree": 0.49024512256128067,
+        "kl": 1.345718805884049,
+        "brier": 0.5798118903688055
       },
       "data/eval_wide.jsonl": {
         "rows": 920,
-        "agree": 0.8402173913043478,
-        "kl": 0.23989550438937327,
-        "brier": 0.09671256435471895
+        "agree": 0.8521739130434782,
+        "kl": 0.24507102074775225,
+        "brier": 0.09771515623017962
       },
       "data/eval_xnli.jsonl": {
         "rows": 740,
-        "agree": 0.8378378378378378,
-        "kl": 0.12820397544657827,
-        "brier": 0.07586474448547421
+        "agree": 0.8405405405405405,
+        "kl": 0.12840145847645495,
+        "brier": 0.07536055317105994
       },
       "data/eval_multi.jsonl": {
         "rows": 1466,
-        "agree": 0.7701227830832197,
-        "kl": 0.18150919490307726,
-        "brier": 0.09427624401568274
+        "agree": 0.7755798090040927,
+        "kl": 0.18438450108898327,
+        "brier": 0.09538443087951312
       }
     }
   },
   "mmlu_pro": {
     "rows": 2000,
-    "accuracy": 0.421
+    "accuracy": 0.4195
   },
   "latency": [
     {
       "questions": 1,
-      "p50_ms": 28.135785833001137,
-      "p99_ms": 30.05402721464634
+      "p50_ms": 27.914823964238167,
+      "p99_ms": 28.93548645079136
     },
     {
       "questions": 10,
-      "p50_ms": 47.89547622203827,
-      "p99_ms": 48.19699190557003
+      "p50_ms": 47.78458643704653,
+      "p99_ms": 48.7666130065918
     }
   ]
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fda2b8f6260d26b60bb6a4561f542f7a7b84a8ef594cb78b8889239df3905b1e
+oid sha256:afcf48bba438429db2abd9d91e8f91752c0a63a970ddd86538d195631c964b13
 size 8411558400

tokenizer_config.json

@@ -9,8 +9,9 @@
   "eos_token": "<|im_end|>",
   "errors": "replace",
   "image_token": "<|image_pad|>",
-  "is_local": false,
+  "is_local": true,
   "local_files_only": false,
+  "max_length": 1024,
   "model_max_length": 262144,
   "model_specific_special_tokens": {
     "audio_bos_token": "<|audio_start|>",
@@ -21,10 +22,16 @@
     "vision_bos_token": "<|vision_start|>",
     "vision_eos_token": "<|vision_end|>"
   },
+  "pad_to_multiple_of": null,
   "pad_token": "<|endoftext|>",
+  "pad_token_type_id": 0,
+  "padding_side": "left",
   "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
   "split_special_tokens": false,
+  "stride": 0,
   "tokenizer_class": "Qwen2Tokenizer",
+  "truncation_side": "left",
+  "truncation_strategy": "longest_first",
   "unk_token": null,
   "video_token": "<|video_pad|>",
   "vision_bos_token": "<|vision_start|>",

train_meta.json

@@ -1,14 +1,14 @@
 {
-  "base": "Qwen/Qwen3.5-4B",
+  "base": "runs/v3",
   "lora": 0,
-  "steps": 11844,
+  "steps": 5922,
   "rows": 378963,
-  "lr": 1e-05,
+  "lr": 5e-06,
   "eval": {
-    "kl_before": 0.3580377897640213,
-    "agree_before": 0.8019111570247934,
-    "kl_after": 0.06131432654503502,
-    "agree_after": 0.9351756198347108
+    "kl_before": 0.06131432654503502,
+    "agree_before": 0.9351756198347108,
+    "kl_after": 0.05286450635965082,
+    "agree_after": 0.9425361570247934
   },
   "data": [
     "data/syn.jsonl",