Model Card — DeepHunted-V4
Model Name: DeepHunted-V4
Base Model: DeepSeek-V4 (fine-tuned)
Training Method: Base -> Supervised Fine-Tuning (SFT) -> GRPO-RL alignment
Domain: Code generation EXCLUSIVELY for Galculator (Calculator/Galculator Interpreter v0.0.1.5)
What's in: Base models and IT models
Overview
DeepHunted-V4 is optimized for generating, completing, and reasoning about code in the custom Galculator language. It emphasizes correctness of expressions, function composition, and environment-aware execution.
Capabilities
- Generate valid Galculator programs (single & multi-line)
- Understand statements vs compound/top-level semantics
- Write/optimize functions (incl. nested, overloads, aliases)
- Handle control expressions (
if,while,for,case) - Manage environments (
env,use,::access) - Apply builtins and logical helpers
- Follow REPL conventions and syntax constraints
Training Data
- Synthetic + curated code datasets in Galculator
- Instruction-response pairs for program synthesis, debugging, and explanation
- RL phase (GRPO) optimized for correctness, syntax validity, and execution fidelity
Limitations
- Narrow specialization: performance outside Galculator may degrade
- Sensitive to strict syntax (e.g.,
let,env, commas in control expressions) - Limited robustness to ambiguous or underspecified prompts
- May produce invalid code if constraints are violated implicitly
Intended Use
- Code generation for Galculator interpreter
- Educational tooling for expression-based languages
- Program synthesis, transformation, and debugging
Out-of-Scope
- General-purpose coding across many languages
- Safety-critical or production-grade execution without validation
Example
let x = 1;
def add(a,b) = a + b;
print(add(x,2));
Evaluation
- Syntax validity rate
- Execution correctness (unit tests)
- RL reward (GRPO objective)
- Pass@k on code generation tasks
Safety
- No built-in sandboxing; execution depends on external interpreter
- Users should validate generated code before running
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support