---
license: gemma
base_model: google/gemma-4-E2B-it
library_name: transformers
pipeline_tag: text-generation
language:
  - en
datasets:
  - build-small-hackathon/slipstream-evm-sft
tags:
  - project-controls
  - earned-value-management
  - forecasting
  - lora
  - distillation
  - code-action
  - agent
  - slipstream
---
# Slipstream - gemma-4-E2B (EVM forecasting agent)

A small **code-action agent** that forecasts a project's final cost (EAC) and finish period from a
mid-flight Earned Value Management snapshot. It is **`google/gemma-4-E2B-it`** (E2B (~2B effective),
Gemma-4 (text decoder)) fine-tuned (LoRA, then merged) to run a single-tool reasoning loop: it writes Python
that calls a curated forecasting toolkit (Earned Schedule, CPI/SPI formulas, a Gompertz growth
curve, a reference-class ML regressor, and the TimesFM / Chronos time-series foundation models),
reconciles their disagreeing estimates, and submits one answer.

It was **distilled from a DeepSeek V4 teacher**: the teacher's reasoning traces over a diverse
simulated project corpus were filtered to a 367-trace SFT set
([`build-small-hackathon/slipstream-evm-sft`](https://huggingface.co/datasets/build-small-hackathon/slipstream-evm-sft)) and the student trained with
assistant-only loss (reasoning + tool-call tokens only). This makes a sub-5B, **edge / air-gapped**
forecaster that matches the classical project-controls baseline and approaches its cloud teacher.

## Results (held-out real projects, 40% complete, n=107)

Scored on 107 real completed projects (Batselier/OR-AS DSLIB), apples-to-apples with every
baseline. `valid` = produced a usable forecast; `EAC error` = median absolute % error on final cost;
`finish error` = median absolute error in periods.

| Method | valid | EAC error | finish error |
| --- | --- | --- | --- |
| **gemma-4-E2B (this model, distilled)** | **0.991** | **2.31%** | **0.63 periods** |
| gemma-4-E2B (base, before distillation) | 0.664 | 3.21% | 0.75 periods |
| Earned Schedule (classical baseline) | 1 | 2.37% | 1 periods |
| DeepSeek V4 teacher (cloud) | 1 | 2.4% | 0.6 periods |

Distillation lifts a base model that could barely operate the tool-call format into a reliable
forecaster that rivals the classical canon and its own teacher.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "build-small-hackathon/slipstream-gemma4-e2b-evm"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="bfloat16")
```

The model is trained to act through a single `run_python(code=...)` tool call and to call
`submit(finish, eac)` from inside that code. See the Slipstream project for the agent loop, the
forecasting toolkit, and the full benchmark.

## Licence and attribution

This is a derivative of **[`google/gemma-4-E2B-it`](https://huggingface.co/google/gemma-4-E2B-it)** and is released
under the **base model's licence** (gemma).
You must comply with the upstream terms.
Training data: [`build-small-hackathon/slipstream-evm-sft`](https://huggingface.co/datasets/build-small-hackathon/slipstream-evm-sft).

Built for the Hugging Face **Build Small** hackathon.