Upload BehzatOne v108 BF16 shards and GGUF artifacts

.gitattributes

@@ -47,3 +47,6 @@ v106/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 gguf/BehzatOne-8B-A1B-v107.BF16.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/BehzatOne-8B-A1B-v107.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
 v107/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+gguf/BehzatOne-8B-A1B-v108.BF16.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+v108/tokenizer.json filter=lfs diff=lfs merge=lfs -text

Modelfile

@@ -1,4 +1,4 @@
-FROM ./gguf/BehzatOne-8B-A1B-v107.Q4_K_M.gguf
+FROM ./gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf
 
 TEMPLATE """<|startoftext|>{{ if .System }}<|im_start|>system
 {{ .System }}<|im_end|>

Modelfile.v108-agent-q4

@@ -0,0 +1,25 @@
+FROM ./gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf
+
+TEMPLATE """<|startoftext|>{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ else }}<|im_start|>system
+You are BehzatOne Agent, a patch-producing coding assistant.
+When asked to edit code, return only a unified git diff. Do not include markdown,
+explanations, <think> blocks, fake tool output, or shell transcripts.<|im_end|>
+{{ end }}<|im_start|>user
+{{ .Prompt }}
+
+Return only the requested artifact. For code edits, return only a unified git diff.<|im_end|>
+<|im_start|>assistant
+"""
+
+RENDERER lfm2-thinking
+PARSER lfm2-thinking
+
+PARAMETER stop "<|im_end|>"
+PARAMETER stop "<|im_start|>"
+PARAMETER temperature 0.1
+PARAMETER top_k 30
+PARAMETER top_p 0.85
+PARAMETER repeat_penalty 1.1
+PARAMETER typical_p 0.9

Modelfile.v108-bf16

@@ -0,0 +1,25 @@
+FROM ./gguf/BehzatOne-8B-A1B-v108.BF16.gguf
+
+TEMPLATE """<|startoftext|>{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ else }}<|im_start|>system
+You are BehzatOne, a careful code assistant by Behzat Industries.
+Answer with the requested final artifact only. Do not include hidden reasoning,
+<think> blocks, or unrelated prose. If you do not know, say UNKNOWN.<|im_end|>
+{{ end }}<|im_start|>user
+{{ .Prompt }}
+
+Answer final only. Do not output <think> blocks.<|im_end|>
+<|im_start|>assistant
+"""
+
+RENDERER lfm2-thinking
+PARSER lfm2-thinking
+
+PARAMETER stop "<|im_end|>"
+PARAMETER stop "<|im_start|>"
+PARAMETER temperature 0.1
+PARAMETER top_k 30
+PARAMETER top_p 0.85
+PARAMETER repeat_penalty 1.1
+PARAMETER typical_p 0.9

Modelfile.v108-q4

@@ -0,0 +1,25 @@
+FROM ./gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf
+
+TEMPLATE """<|startoftext|>{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ else }}<|im_start|>system
+You are BehzatOne, a careful code assistant by Behzat Industries.
+Answer with the requested final artifact only. Do not include hidden reasoning,
+<think> blocks, or unrelated prose. If you do not know, say UNKNOWN.<|im_end|>
+{{ end }}<|im_start|>user
+{{ .Prompt }}
+
+Answer final only. Do not output <think> blocks.<|im_end|>
+<|im_start|>assistant
+"""
+
+RENDERER lfm2-thinking
+PARSER lfm2-thinking
+
+PARAMETER stop "<|im_end|>"
+PARAMETER stop "<|im_start|>"
+PARAMETER temperature 0.1
+PARAMETER top_k 30
+PARAMETER top_p 0.85
+PARAMETER repeat_penalty 1.1
+PARAMETER typical_p 0.9

README.md

@@ -17,35 +17,33 @@ language:
 pipeline_tag: text-generation
 ---
 
-# BehzatOne-8B-A1B v107
+# BehzatOne-8B-A1B v108
 
-This release builds on the v105 BF16 base and merges the current v106 SFT and DPO adapters into it. It ships the merged BF16 weights plus BF16 and Q4_K_M GGUF artifacts for local deployment.
+This is a targeted repair release built on v107. It focuses on strict agentic output contracts: valid unified diffs, complete file rewrites, JSON-only answers, and rejecting tool-schema regurgitation or fake shell transcripts.
 
 ## Version Lineage
 
 - v105: ~109k-sample release with SFT plus anti-hallucination DPO.
 - v106: agentic SFT plus DPO release.
 - v107: merge release built from the v105 BF16 base plus the current v106 SFT and DPO adapters.
+- v108: targeted agentic-format repair release on top of v107, focused on exact patch/file/JSON output and anti-schema-regurgitation.
 
 
 ## What changed
 
-- New SFT adapter merged into the base `LiquidAI/LFM2.5-8B-A1B`
-- DPO retrained with a public 80/10/10 mix:
-  - 80% agentic/tool-use preferences
-  - 10% coding preferences
-  - 10% conversation preferences
-- GGUF artifacts generated for both BF16 and Q4_K_M
-- Ollama Modelfiles included for BF16 and Q4 deployment
+- Merged BF16 weights for this release.
+- GGUF artifacts generated for both BF16 and Q4_K_M.
+- Ollama Modelfiles included for BF16 and Q4 deployment.
+- See the release summary above for the training focus of this version.
 
 ## Artifacts
 
-- `/v107/` - merged BF16 HF model shards
-- `gguf/BehzatOne-8B-A1B-v107.BF16.gguf` - BF16 GGUF
-- `gguf/BehzatOne-8B-A1B-v107.Q4_K_M.gguf` - Q4_K_M GGUF
-- `Modelfile.v107-bf16`
-- `Modelfile.v107-q4`
-- `Modelfile.v107-agent-q4`
+- `/v108/` - merged BF16 HF model shards
+- `gguf/BehzatOne-8B-A1B-v108.BF16.gguf` - BF16 GGUF
+- `gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf` - Q4_K_M GGUF
+- `Modelfile.v108-bf16`
+- `Modelfile.v108-q4`
+- `Modelfile.v108-agent-q4`
 
 ## Status
 
@@ -61,13 +59,13 @@ import torch
 
 model = AutoModelForCausalLM.from_pretrained(
     "behzatindustries/BehzatOne-8B-A1B",
-    subfolder="v107",
+    subfolder="v108",
     torch_dtype=torch.bfloat16,
     trust_remote_code=True,
 ).cuda()
 tokenizer = AutoTokenizer.from_pretrained(
     "behzatindustries/BehzatOne-8B-A1B",
-    subfolder="v107",
+    subfolder="v108",
     trust_remote_code=True,
 )
 ```

gguf/BehzatOne-8B-A1B-v108.BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d51cb0ecb9d088b613be2deed2b93cd6e02c0cfeea69c3833c39fe1b7df9217
+size 16947255424

gguf/BehzatOne-8B-A1B-v108.Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d91d7d5e680f2dee608dcd0f785db0ecc7d78027bae981e93075f27f3b3d079
+size 5155559552

v108/config.json

@@ -0,0 +1,61 @@
+{
+  "architectures": [
+    "Lfm2MoeForCausalLM"
+  ],
+  "bos_token_id": 124894,
+  "conv_L_cache": 3,
+  "conv_bias": false,
+  "dtype": "bfloat16",
+  "eos_token_id": 124900,
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 7168,
+  "layer_types": [
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv",
+    "full_attention",
+    "conv",
+    "conv"
+  ],
+  "max_position_embeddings": 128000,
+  "model_type": "lfm2_moe",
+  "moe_intermediate_size": 1792,
+  "norm_eps": 1e-05,
+  "norm_topk_prob": true,
+  "num_attention_heads": 32,
+  "num_dense_layers": 2,
+  "num_experts": 32,
+  "num_experts_per_tok": 4,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 8,
+  "pad_token_id": 124893,
+  "rope_parameters": {
+    "rope_theta": 5000000,
+    "rope_type": "default"
+  },
+  "routed_scaling_factor": 1.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.12.1",
+  "use_cache": true,
+  "use_expert_bias": true,
+  "vocab_size": 128000
+}

v108/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 124894,
+  "do_sample": true,
+  "eos_token_id": 124900,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 124893,
+  "repetition_penalty": 1.05,
+  "temperature": 0.2,
+  "top_k": 80,
+  "transformers_version": "5.12.1",
+  "use_cache": true
+}

v108/model-00001-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8c6f7f807b680fac78a1b17e6b99fbc69761fdf179a4ef489b3108dbc20612c
+size 3943343776

v108/model-00002-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c52129cf229ab39615a7a025fa43fb13ce5f426b059822dbe638a2db1d1edf82
+size 3700188296

v108/model-00003-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c41f2c4348e014ad372039ce2a09c33c32d64cb89c0adbed488692e131307f3a
+size 3679216200

v108/model-00004-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf6bb60d5672a46fcf73589080f2273af7299a8e4581215c42c42045ce1e2964
+size 3645649144

v108/model-00005-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2db98668c94632eecad3ef946be1eae1d712923f4a33d1efc15d8f427c38ef95
+size 1967606440

v108/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:695be7802a0e4b8a81048f0ff5ebb7fc811a0ba5a6be63dbb24deb5a81096f41
+size 17905598

v108/tokenizer_config.json

@@ -0,0 +1,12 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|startoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<|pad|>",
+  "tokenizer_class": "TokenizersBackend",
+  "use_default_system_prompt": false
+}