File size: 2,336 Bytes
e93c701 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | ---
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.
|