README.md

---
license: mit 
language:
- en
base_model: Qwen/Qwen3-1.7B
tags:
- reasoning
- math
- code
- grpo
- reinforcement-learning
- Atomight V2.5
pipeline_tag: text-generation
---

# Atomight-V2.5-1.7B-C1

<p align="center">
  <img src="OfficialAtomight.png" alt="Atomight Logo" width="500" style="max-width: 100%;">
</p>

**Atomight V2.5 1.7B**, our *most capable and intelligent model* yet, is a multi-domain reasoning model built on **Qwen3-1.7B**, trained with **GRPO** (Group Relative Policy Optimization) across STEM, science, math, and code, using a curated six-dataset stack in a streaming round-robin setup — trained entirely on a single Colab free-tier T4 GPU.
**Note: This is only the first checkpoint model of Atomight-V2.5-1.7B, but still usable and powerful model.**
## Overview

Atomight-V2.5 is post-trained from `Qwen/Qwen3-1.7B` (already instruct/chat-capable) using memory-efficient chunked GRPO to fit T4 VRAM constraints. Training data spans four domains, each selected for having genuinely verifiable rewards (exact-match answers or executable test cases) rather than relying purely on LLM-judged quality:

| Domain | Dataset(s) | Role |
|---|---|---|
| STEM | `Logics-MLLM/Logics-STEM-SFT-Dataset-Open-5.3M` | Broad STEM coverage, difficulty-annotated |
| Science | `MegaScience/MegaScience` | Science-specific depth, ablation-selected subsets |
| Math | `nvidia/OpenMathReasoning`, `zwhe99/DeepMath-103K` | Verifiable-answer CoT and RL-native reward data |
| Code | `nvidia/OpenCodeInstruct` | Instruction-tuned code generation with test cases |

Reward shaping combines an XML-format reward (`<thinking>` / `<answer>` tag structure) with an exact-match correctness reward, evaluated via `trl`'s `GRPOTrainer` on 4-bit QLoRA (Unsloth).

## Benchmark Results

Evaluated with [EleutherAI's `lm-evaluation-harness`](https://github.com/EleutherAI/lm-evaluation-harness), standard few-shot settings, no sample limiting.

| Benchmark | Score |
|---|---|
| GSM8K (strict-match) | **69.6%** |
| GSM8K (flexible-extract) | **69.9%** |
| MMLU (overall) | **55.7%** |
| ARC-Challenge (acc_norm) | **43.0%** |
| HellaSwag (acc_norm) | **60.4%** |
| Winogrande | **61.1%** |
| TruthfulQA (mc2) | **45.9%** |
| HumanEval (pass@1) | **40.2%** |
| MBPP (pass@1) | **42.8%** |

### Full benchmark profile

![Atomight-V2.5-1.7B full benchmark radar](./assets/atomight_radar_full_profile.jpg) 

### MMLU breakdown

| Category | Score |
|---|---|
| STEM | 53.9% |
| Social Sciences | 63.8% |
| Humanities | 48.7% |
| Other | 60.0% |

## Progress vs. Atomight-V2.1-0.5B

V2.5 (1.7B) is compared here against the previous-generation V2.1 (0.5B) on the benchmarks both models were evaluated on. Note this comparison reflects both the parameter increase (0.5B → 1.7B) and the change in training methodology — it isn't an isolated ablation of either factor alone.

![V2.1 vs V2.5 radar comparison](./assets/atomight_radar_v21_vs_v25.png)

![V2.1 vs V2.5 bar comparison](./assets/atomight_bar_v21_vs_v25.png)

| Benchmark | V2.1-0.5B | V2.5-1.7B | Δ |
|---|---|---|---|
| GSM8K (strict) | 19.8% | 69.6% | **+49.8** |
| GSM8K (flexible) | 32.4% | 69.9% | **+37.5** |
| ARC-Challenge | 33.8% | 43.0% | **+9.2** |
| HellaSwag | 52.4% | 60.4% | **+8.0** |

Notably, the gap between GSM8K's strict-match and flexible-extract scores nearly closed between versions (12.6-point gap in V2.1 → 0.3-point gap in V2.5), indicating the GRPO reward shaping improved not just raw math ability but the model's consistency in producing cleanly parseable, correctly formatted answers.

## Known Limitations

**Word-problem math vs. abstract math.** Despite strong GSM8K performance, Atomight-V2.5 underperforms on MMLU's more abstract/formal math and physics subcategories:

![Math transfer gap](./assets/atomight_math_transfer_gap.png)

| Category | Score |
|---|---|
| GSM8K (word problems) | 69.6% |
| MMLU College Mathematics | 38.0% |
| MMLU High School Mathematics | 35.9% |
| MMLU College Physics | 34.3% |
| MMLU Abstract Algebra | 39.0% |

This suggests the training data (competition/olympiad-style word problems) taught strong step-by-step arithmetic reasoning, but that skill did not fully transfer to more abstract, formally-structured mathematics.

**Truthfulness.** TruthfulQA (45.9%) was not a training target — the model has no specific resistance to plausible-but-false completions beyond what emerges from base pretraining.

**Code generalization.** HumanEval/MBPP scores (40-43%) are solid for model size but sit below the model's math performance in relative terms, suggesting a similar narrow-transfer pattern to the math gap above.

## Training Infrastructure

- **Base model:** `Qwen/Qwen3-1.7B` (post-trained, not a base checkpoint)
- **Method:** GRPO via `trl`, 4-bit QLoRA via Unsloth (`r=16`, `lora_alpha=32`)
- **Hardware:** Single Google Colab free-tier Tesla T4 (16GB)
- **Memory strategy:** Streamed, chunked micro-batches (20 samples/chunk) pulled round-robin across all six datasets, trained, then discarded — avoids materializing multi-million-row datasets in memory
- **Reward functions:** XML-format adherence (`<thinking>`/`<answer>` tags) + exact-match correctness against verifiable answers

## Evaluation Methodology

All scores use standard `lm-evaluation-harness` task configs at default few-shot settings, with no `--limit` sampling — results are intended to be directly comparable to other models' published numbers on the same tasks. Code benchmarks (HumanEval, MBPP) were run with `HF_ALLOW_CODE_EVAL=1` in a sandboxed Colab runtime.

## Citation

If you use this model, please cite the underlying datasets and base model:

```
Base model: Qwen/Qwen3-1.7B (Qwen Team, Alibaba)
Training datasets: Logics-STEM-SFT-Dataset-Open-5.3M, MegaScience,
OpenMathReasoning, DeepMath-103K, OpenCodeInstruct
```