Add exp8_soyuz: EAGLE draft fine-tuned on Soyuz-sft clean (agentic SWE), tau=5.31

experiments/exp8_soyuz/NOTE.md

@@ -0,0 +1,8 @@
+# exp8_soyuz — EAGLE-3 draft fine-tuned on Soyuz-sft clean (agentic SWE)
+Continue-trained from exp7 (champion) on AlexWortega/Soyuz-sft `clean/` agentic SWE trajectories.
+- gen: max_length 8192, aux layers [1,23,45], 250 train / 60 eval samples.
+- train: 10 epochs, lr 1e-4, resumed exp7 weights.
+- held-out (Soyuz agentic eval, 60 samples): tau=5.31, acc0=0.78, per-pos 0.78/0.77/0.76/0.76/0.75/0.75/0.74.
+- NOTE: tau measured on Soyuz agentic eval (not the code eval used for exp7 tau=4.25); the flat high
+  acceptance reflects the high predictability of agentic SWE traces (tool-call boilerplate, paths).
+  This draft targets North-Mini-Code's actual agentic-coding usage domain.

experiments/exp8_soyuz/config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "LlamaForCausalLMEagle3"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 2,
+  "draft_vocab_size": 32000,
+  "dtype": "bfloat16",
+  "eos_token_id": 255001,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "max_position_embeddings": 4096,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 1,
+  "num_key_value_heads": 4,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 50000.0,
+  "target_hidden_size": 2048,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.1",
+  "use_cache": true,
+  "vocab_size": 262144
+}

experiments/exp8_soyuz/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a721ed9b80e667d3b92b5806d6efbb29aca36d0383a58240342eb061bd747369
+size 366489056