jersobh/ jeff-2b-decision:latest

21 1 week ago

Fine-tuned on top of Qwen/Qwen3.5-2B-Base with extended 5-epoch training and hard-negative mining. Optimized for structured schema-guided routing and intent classification.

vision tools decision thinking 
curl http://localhost:11434/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jersobh/jeff-2b-decision",
    "state": "Hello World",
    "questions": {
      "says_hello": {
        "type": "noul",
        "instructions": "Does the state text contain a greeting?",
        "criteria": {
          "true": "The state text contains a greeting.",
          "false": "The state text does not contain a greeting."
        }
      }
    }
  }'

Details

1 week ago

5421fa6d3714 · 1.3GB

model
qwen35
·
1.88B
·
Q4_K_M
template
{{ if .System }}{{ .System }}{{ end }} Task: {{ .Prompt }} Selection:
system
You are a strict JSON decision engine. Match the task to the correct tool index from the choices. Ou
params
{ "num_ctx": 4096, "repeat_penalty": 1.2, "temperature": 0, "top_p": 0.7 }

Readme

language: - en license: apache-2.0 tags: - unsloth - qlora - gguf - decision-engine - routing

base_model: internlm/Intern-Decision-2B

Jeff-2B Decision Model (Optimized Dual Precision Benchmarks)

Fine-tuned on top of internlm/Intern-Decision-2B using Unsloth with extended 5-epoch training and hard-negative mining. Optimized for structured schema-guided routing and intent classification.

Benchmark Results Comparison

Benchmark / Task Unquantized (BF16) Quantized (Q4_K_M GGUF)
BANKING77 Intent Classification 95.80% 95.60%
MASSIVE Intent Classification 87.10% 87.00%
CLINC150 Intent Classification 95.20% 95.00%
Type-Decision Routing (Dolly) 90.30% 90.10%

Training Configuration

  • Base Model: internlm/Intern-Decision-2B
  • Precision: Native BF16 (NVIDIA L4 24GB)
  • Method: LoRA (Rank: 32, Alpha: 32)
  • Epochs: 5 (Hard-negative mining enabled)
  • Export Precision: 4-bit Quantized GGUF (Q4_K_M)

Usage with ollama

ollama run jersobh/jeff-2b-decision --format json "Search the knowledge graph for customer interaction logs regarding subscription renewals. Choices: 1. SQLDatabaseQueryTool, 2. VectorSearchTool, 3. WebScraperTool, 4. ChatDirectly"