| --- |
| license: apache-2.0 |
| base_model: thinkingmachines/Inkling |
| library_name: peft |
| pipeline_tag: text-generation |
| language: |
| - en |
| tags: |
| - finance |
| - trading |
| - lora |
| - tinker |
| - inkling |
| --- |
| |
| # TradeFinder 1 |
|
|
| TradeFinder 1 is an AI model by [RobinArena](https://github.com/RobinArena). |
| It is a LoRA adapter for |
| [`thinkingmachines/Inkling`](https://huggingface.co/thinkingmachines/Inkling), |
| trained with the Thinking Machines Tinker API. |
|
|
| The model was trained to produce structured stock-trading decisions and to |
| review prior decisions after their realized outcomes are known. Review examples |
| include the original decision, subsequent return, and position PnL. Targets |
| identify decisions as right or wrong, correct faulty reasoning, and preserve |
| useful reasoning from successful decisions. |
|
|
| ## Files |
|
|
| `adapter_model.safetensors` contains the trained LoRA weights. |
| `adapter_config.json` records the PEFT configuration and Inkling base model. |
| This repository contains an adapter, so the Inkling base weights are required |
| for local inference. |
|
|
| Training examples combined past-only daily market features, portfolio state, |
| structured buy, hold, and sell decisions, retrospective outcome reviews, and |
| explicit model-identity examples. Future returns and PnL were used only in |
| retrospective labels and were excluded from live decision prompts. |
|
|
| ## Inference |
|
|
| The adapter was trained with Tinker's `tml_v0` renderer. The matching hosted |
| sampler is the most direct way to reproduce training-time rendering. Local PEFT, |
| llama.cpp, vLLM, and SGLang support depends on their Inkling adapter support. |
| At publication time, llama.cpp can run quantized Inkling base models but its |
| LoRA-to-GGUF converter does not support `InklingForConditionalGeneration`. |
|
|
| ## Intended use |
|
|
| TradeFinder 1 is intended for research, model evaluation, and paper trading. |
| Its outputs can be incorrect, stale, or poorly calibrated. It does not provide |
| personalized financial advice. Validate decisions against current primary data, |
| apply portfolio-level risk controls, and require human approval before placing |
| trades. Historical correctness and low training loss do not establish future |
| profitability. |
|
|
| ## Attribution |
|
|
| Developed and fine-tuned by RobinArena. The base Inkling model is published by |
| Thinking Machines Lab under its stated license and acceptable-use terms. |
|
|
