Upload Deco-G HMM checkpoints

README.md

@@ -0,0 +1,21 @@
+---
+tags:
+- deco-g
+- hmm
+- constrained-decoding
+- format-estimation
+---
+
+# Deco-G HMM Checkpoints
+
+This repository hosts the HMM checkpoints used by Deco-G for the public GSM8K and SummEval experiments in [Decoupling Task-Solving and Output Formatting in LLM Generation](https://arxiv.org/abs/2510.03595).
+
+The code release at https://github.com/haikangdeng/deco-g downloads these checkpoints automatically for Deco-G decoding. The checkpoints are organized by model-specific subfolder:
+
+| Deco-G alias | Base model | Subfolder |
+| --- | --- | --- |
+| `llama8b` | `meta-llama/Llama-3.1-8B-Instruct` | `llama3.1-8b-instruct` |
+| `qwen7b` | `Qwen/Qwen2.5-7B-Instruct` | `qwen2.5-7b-instruct` |
+| `qwen3-8b` | `Qwen/Qwen3-8B` | `qwen3-8b` |
+
+Each subfolder contains a `config.json` and `model.safetensors` compatible with Deco-G's `HMM` class.

llama3.1-8b-instruct/README.md

@@ -0,0 +1,20 @@
+---
+tags:
+- deco-g
+- hmm
+- constrained-decoding
+- format-estimation
+base_model: meta-llama/Llama-3.1-8B-Instruct
+---
+
+# Deco-G HMM: meta-llama/Llama-3.1-8B-Instruct
+
+This subfolder contains the Format Estimation Module HMM checkpoint for `llama8b` in Deco-G.
+
+- Paper: [Decoupling Task-Solving and Output Formatting in LLM Generation](https://arxiv.org/abs/2510.03595)
+- Base model: `meta-llama/Llama-3.1-8B-Instruct`
+- Hidden states: `4096`
+- Vocabulary size: `128256`
+- EOS token id: `128009`
+
+The checkpoint is loaded automatically by the Deco-G release code when running with `--model llama8b --control_type decog`.

llama3.1-8b-instruct/config.json

@@ -0,0 +1,5 @@
+{
+  "eos_token_id": 128009,
+  "hidden_states": 4096,
+  "vocab_size": 128256
+}

llama3.1-8b-instruct/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bafca45da4a3c758bad1201b855b388b39213fe88c83eef77a40408055c9e52b
+size 2168471800

manifest.json

@@ -0,0 +1,31 @@
+{
+  "repo": "hk/decog-hmms",
+  "paper": "Decoupling Task-Solving and Output Formatting in LLM Generation",
+  "arxiv": "https://arxiv.org/abs/2510.03595",
+  "checkpoints": [
+    {
+      "model_alias": "llama8b",
+      "subfolder": "llama3.1-8b-instruct",
+      "base_model": "meta-llama/Llama-3.1-8B-Instruct",
+      "hidden_states": 4096,
+      "vocab_size": 128256,
+      "eos_token_id": 128009
+    },
+    {
+      "model_alias": "qwen7b",
+      "subfolder": "qwen2.5-7b-instruct",
+      "base_model": "Qwen/Qwen2.5-7B-Instruct",
+      "hidden_states": 4096,
+      "vocab_size": 152064,
+      "eos_token_id": 151645
+    },
+    {
+      "model_alias": "qwen3-8b",
+      "subfolder": "qwen3-8b",
+      "base_model": "Qwen/Qwen3-8B",
+      "hidden_states": 4096,
+      "vocab_size": 151936,
+      "eos_token_id": 151645
+    }
+  ]
+}

qwen2.5-7b-instruct/README.md

@@ -0,0 +1,20 @@
+---
+tags:
+- deco-g
+- hmm
+- constrained-decoding
+- format-estimation
+base_model: Qwen/Qwen2.5-7B-Instruct
+---
+
+# Deco-G HMM: Qwen/Qwen2.5-7B-Instruct
+
+This subfolder contains the Format Estimation Module HMM checkpoint for `qwen7b` in Deco-G.
+
+- Paper: [Decoupling Task-Solving and Output Formatting in LLM Generation](https://arxiv.org/abs/2510.03595)
+- Base model: `Qwen/Qwen2.5-7B-Instruct`
+- Hidden states: `4096`
+- Vocabulary size: `152064`
+- EOS token id: `151645`
+
+The checkpoint is loaded automatically by the Deco-G release code when running with `--model qwen7b --control_type decog`.

qwen2.5-7b-instruct/config.json

@@ -0,0 +1,5 @@
+{
+  "eos_token_id": 151645,
+  "hidden_states": 4096,
+  "vocab_size": 152064
+}

qwen2.5-7b-instruct/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1223a4107febf59a31cfb0c27db40fafec6a003ec240c67710da9ddf41e642a
+size 2558542072

qwen3-8b/README.md

@@ -0,0 +1,20 @@
+---
+tags:
+- deco-g
+- hmm
+- constrained-decoding
+- format-estimation
+base_model: Qwen/Qwen3-8B
+---
+
+# Deco-G HMM: Qwen/Qwen3-8B
+
+This subfolder contains the Format Estimation Module HMM checkpoint for `qwen3-8b` in Deco-G.
+
+- Paper: [Decoupling Task-Solving and Output Formatting in LLM Generation](https://arxiv.org/abs/2510.03595)
+- Base model: `Qwen/Qwen3-8B`
+- Hidden states: `4096`
+- Vocabulary size: `151936`
+- EOS token id: `151645`
+
+The checkpoint is loaded automatically by the Deco-G release code when running with `--model qwen3-8b --control_type decog`.

qwen3-8b/config.json

@@ -0,0 +1,5 @@
+{
+  "eos_token_id": 151645,
+  "hidden_states": 4096,
+  "vocab_size": 151936
+}

qwen3-8b/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1877c11981c34685d851908844d1f17c0c5fc39a66ab54c20687674d5c7f76c3
+size 2556444920