358 lines
17 KiB
Markdown
358 lines
17 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: Qwen/Qwen3-0.6B
|
||
tags:
|
||
- unsloth
|
||
- lora
|
||
- qwen3
|
||
- reasoning
|
||
- distillation
|
||
- conversational
|
||
datasets:
|
||
- EphAsad/QWENMillenium-SF
|
||
- EphAsad/Phi4Millennium-SF
|
||
- EphAsad/MistralMillenium-SF
|
||
- Modotte/CodeX-2M-Thinking
|
||
- Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned
|
||
- WithinUsAI/MiniMax_M2.7_Distilled_5k
|
||
- tuanha1305/DeepSeek-R1-Distill
|
||
- open-r1/OpenThoughts-114k-math
|
||
- flytech/python-codes-25k
|
||
- FreedomIntelligence/medical-o1-reasoning-SFT
|
||
- Jackrong/Claude-opus-4.7-TraceInversion-5000x
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
---
|
||
|
||
|
||
|
||
# Atem-0.6B
|
||
|
||
*Ancient logic. Modern intelligence.*
|
||
|
||
A 0.6B reasoning model trained via multi-source knowledge distillation from frontier teacher models.
|
||
|
||
|
||
|
||
---
|
||
|
||
## Overview
|
||
|
||
Atem-0.6B is a 0.6B parameter reasoning model built via supervised fine-tuning on a curated corpus of approximately 120,000 examples distilled from multiple frontier teacher models. Starting from Qwen/Qwen3-0.6B, Atem was trained using LoRA to preserve base model capabilities while shifting output style toward clean, directly-formatted final answers.
|
||
|
||
This is **Stage 1** of a planned multi-stage training series, and the first entry in the Atem family built on Qwen3 rather than Qwen2.5. Stage 1 strips `<think>` reasoning traces from all training data, deliberately suppressing Qwen3's native exposed chain-of-thought in favor of direct answers. **Stage 2 (Atem-Savant-0.6B) is currently in progress**, layering curated chain-of-thought traces back on top of this foundation — see [Known Limitations](#known-limitations) for why that stage matters.
|
||
|
||
---
|
||
|
||
## Model Details
|
||
|
||
| Property | Value |
|
||
| ------------------------ | ---------------------------------------- |
|
||
| **Base model** | Qwen/Qwen3-0.6B |
|
||
| **Training method** | LoRA Supervised Fine-Tuning (Stage 1) |
|
||
| **LoRA config** | r=32, alpha=64, dropout=0.05 |
|
||
| **Target modules** | q, k, v, o, gate, up, down projections |
|
||
| **Parameters** | ~596M |
|
||
| **Trainable (LoRA) params** | 20,185,088 (3.28% of base) |
|
||
| **Training records** | 120,017 |
|
||
| **Epochs** | 2 |
|
||
| **Effective batch size** | 128 (batch 32 × grad accum 4) |
|
||
| **Learning rate** | 2e-4, cosine schedule, 5% warmup |
|
||
| **Final train loss** | 1.055 |
|
||
| **Final val loss** | 1.073 |
|
||
| **Hardware** | NVIDIA A100-SXM4 80GB |
|
||
| **Max sequence length** | 4,096 tokens |
|
||
| **Precision** | bfloat16 |
|
||
| **License** | Apache 2.0 |
|
||
|
||
---
|
||
|
||
## Intended Use
|
||
|
||
Atem-0.6B is designed for lightweight, open-ended reasoning tasks where structured, direct answers add value at low compute cost:
|
||
|
||
- Code explanation, implementation, and debugging
|
||
- Mathematical problem solving with working shown
|
||
- Analytical reasoning and hypothesis evaluation
|
||
- Concept explanation and comparative analysis
|
||
- Logic, argument, and fallacy identification
|
||
|
||
Atem-0.6B is **not** designed for retrieval-heavy factual lookup, real-time information, or tasks requiring broad knowledge breadth beyond its training domains. At 0.6B parameters its capability ceiling is naturally lower than larger Atem models — expect it to be most useful where speed and footprint matter more than depth on hard, multi-step problems.
|
||
|
||
---
|
||
|
||
## Training Data
|
||
|
||
Atem-0.6B was trained on a corpus assembled from eleven sources, combining domain-specific generated datasets and publicly available distillation datasets from frontier models. All outputs containing `<think>` reasoning traces were stripped to clean final responses for Stage 1 training.
|
||
|
||
| Dataset | Records | Source / Teacher |
|
||
| --------------------------------------------- | ------------ | ----------------------------------------------------- |
|
||
| EphAsad/QWENMillenium-SF | ~ | Qwen2.5-14B — Analytical & Scientific |
|
||
| EphAsad/Phi4Millennium-SF | ~ | Phi-4 14B — Mathematical Reasoning |
|
||
| EphAsad/MistralMillenium-SF | ~ | Mistral-Nemo-12B — Language & Comprehension |
|
||
| Modotte/CodeX-2M-Thinking | 40,000 | Mixed — Coding |
|
||
| Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned | 23,000 | Kimi K2.5 — General Distillation (English filtered) |
|
||
| WithinUsAI/MiniMax_M2.7_Distilled_5k | 5,000 | MiniMax M2.7 |
|
||
| tuanha1305/DeepSeek-R1-Distill | 9,000 | DeepSeek-R1 |
|
||
| open-r1/OpenThoughts-114k-math | 10,000 | Mixed — Mathematics (correct answers only) |
|
||
| flytech/python-codes-25k | 10,000 | Python coding |
|
||
| FreedomIntelligence/medical-o1-reasoning-SFT | 10,000 | Medical reasoning (English config) |
|
||
| Jackrong/Claude-opus-4.7-TraceInversion-5000x | 5,000 | Claude Opus 4.7 — Trace Inversion |
|
||
| **Total** | **120,017** | |
|
||
|
||
---
|
||
|
||
## Training Configuration
|
||
|
||
```python
|
||
# Key hyperparameters
|
||
lora_r = 32
|
||
lora_alpha = 64
|
||
lora_dropout = 0.05
|
||
max_seq_length = 4096
|
||
learning_rate = 2e-4
|
||
lr_scheduler = 'cosine'
|
||
warmup_ratio = 0.05
|
||
batch_size = 32
|
||
grad_accumulation = 4 # effective batch size: 128
|
||
num_epochs = 2
|
||
dtype = bfloat16
|
||
load_in_4bit = True # during training
|
||
|
||
```
|
||
|
||
Training used Unsloth with `train_on_responses_only` masking, ensuring loss was computed exclusively on assistant response tokens. Because Qwen3 ships with no default system prompt (unlike Qwen2.5-Instruct), Atem's identity is baked in via a chat-template modification that injects Atem as the default persona only when no explicit system message is supplied — explicit system messages still take priority. A pre-training validation suite verified this injection, confirmed the response-masking boundary correctly accounts for Qwen3's automatic empty `<think></think>` scaffold insertion, and checked for leaked reasoning content before training began.
|
||
|
||
After training, LoRA adapters were merged into the base weights and exported as a full merged model.
|
||
|
||
**Loss curve:**
|
||
|
||
| Step | Train Loss | Val Loss |
|
||
| ----- | ---------- | --------- |
|
||
| 200 | 1.166 | 1.163 |
|
||
| 800 | 1.108 | 1.096 |
|
||
| 1400 | 0.983 | 1.077 |
|
||
| Final (1876) | **1.055** | **1.073** |
|
||
|
||
Validation loss plateaued around step 1600 of 1876 total steps — the final ~15% of training produced only marginal further improvement (1.074 → 1.073). Train loss showed some batch-to-batch volatility late in training (a step-1800 spike to 1.088, consistent with the dataset's domain diversity rather than divergence), but validation loss stayed smooth and never reversed, indicating no overfitting across the two epochs.
|
||
|
||
---
|
||
|
||
## Evaluation
|
||
|
||
### Benchmark Results
|
||
|
||
Evaluated against the base model (`unsloth/qwen3-0.6b-unsloth-bnb-4bit`) using lm-evaluation-harness.
|
||
|
||
| Task | Base (Qwen3-0.6B) | Atem-0.6B | Delta |
|
||
| ------------------------ | ------------------ | ---------- | ----------- |
|
||
| ARC-Challenge (0-shot, acc_norm) | 33.0% | 35.0% | +2.0% ✓ |
|
||
| GSM8K (5-shot, strict-match) | 26.7% | **31.8%** | **+5.1%** ✓ |
|
||
| HellaSwag (0-shot, acc_norm) | 45.3% | 45.8% | +0.5% |
|
||
|
||
**Eval condition note:** the base model was loaded in 4-bit (`unsloth/qwen3-0.6b-unsloth-bnb-4bit`); Atem-0.6B was evaluated as the full bfloat16 merged model. This is not a precision-matched comparison — the gap may be modestly inflated relative to a 4-bit-vs-4-bit or bf16-vs-bf16 run. GSM8K used 5-shot prompting per lm-eval's default config; ARC-Challenge and HellaSwag were 0-shot.
|
||
|
||
The GSM8K gain is the standout figure, but it likely reflects Stage 1's training toward clean, directly-formatted final answers — which matters a great deal for lm-eval's exact-match-on-extracted-number scoring — more than a deeper improvement in multi-step mathematical reasoning. The qualitative evaluation below, which looks at harder, less templated problems, supports this reading: reasoning depth on multi-step problems is not uniformly better than the base model. ARC-Challenge and HellaSwag, which probe general/commonsense knowledge rather than output formatting, moved only slightly — expected, since SFT on this corpus isn't designed to add new general knowledge.
|
||
|
||
### Qualitative Evaluation
|
||
|
||
Atem-0.6B was evaluated against base Qwen3-0.6B (default thinking-enabled) across 30 domain-representative questions with matched system prompts.
|
||
|
||
| Domain | Questions | Outcome |
|
||
| --------------------- | --------- | ----------------------------------------------------------------------------------------- |
|
||
| Coding | 8 | Mixed — comparable correctness; Atem notably more concise and direct |
|
||
| Mathematics | 6 | Mixed — base model's exposed reasoning self-corrects mid-generation on some multi-step problems that Atem commits to an error on |
|
||
| Analytical Reasoning | 6 | Base model edges ahead — exposed reasoning gives more room to work through multi-step arguments |
|
||
| General Knowledge | 5 | Comparable |
|
||
| Language & Logic | 5 | Comparable, slight edge to base on illustrative examples |
|
||
|
||
Atem-0.6B's outputs were consistently more concise and directly formatted — a direct result of Stage 1's design goal of suppressing exposed chain-of-thought. This did **not** translate into a uniform quality advantage over the base model: on problems requiring several sequential reasoning steps, the base model's visible thinking trace sometimes catches and corrects mistakes mid-generation that Atem, having no scratchpad, does not. This is the expected cost of the no-think format rather than a knowledge regression, and is the explicit target of the in-progress Stage 2 (Atem-Savant-0.6B) training, which reintroduces chain-of-thought.
|
||
|
||
One output during qualitative testing showed repetitive/degenerate text (a duplicated bullet list) on a single open-ended analytical question — noted here for transparency rather than treated as representative.
|
||
|
||
---
|
||
|
||
## Usage
|
||
|
||
### Transformers
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
import torch
|
||
|
||
model_name = "EphAsad/Atem-0.6B"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype=torch.bfloat16,
|
||
device_map="auto"
|
||
)
|
||
|
||
messages = [
|
||
{
|
||
"role": "user",
|
||
"content": "Write a Python function that checks whether a number is prime."
|
||
}
|
||
]
|
||
|
||
inputs = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=True,
|
||
add_generation_prompt=True,
|
||
return_tensors="pt"
|
||
).to(model.device)
|
||
|
||
with torch.no_grad():
|
||
output = model.generate(
|
||
input_ids=inputs,
|
||
max_new_tokens=1000,
|
||
temperature=0.7,
|
||
top_p=0.9,
|
||
repetition_penalty=1.1,
|
||
do_sample=True,
|
||
)
|
||
|
||
response = tokenizer.decode(
|
||
output[0][inputs.shape[1]:],
|
||
skip_special_tokens=True
|
||
)
|
||
print(response)
|
||
|
||
```
|
||
|
||
### Unsloth (faster inference)
|
||
|
||
```python
|
||
from unsloth import FastLanguageModel
|
||
import torch
|
||
|
||
model, tokenizer = FastLanguageModel.from_pretrained(
|
||
model_name="EphAsad/Atem-0.6B",
|
||
max_seq_length=4096,
|
||
dtype=torch.bfloat16,
|
||
load_in_4bit=True,
|
||
)
|
||
FastLanguageModel.for_inference(model)
|
||
|
||
messages = [
|
||
{
|
||
"role": "user",
|
||
"content": "Explain the difference between a stack and a queue, with examples."
|
||
}
|
||
]
|
||
|
||
inputs = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=True,
|
||
add_generation_prompt=True,
|
||
return_tensors="pt"
|
||
).to("cuda")
|
||
|
||
with torch.no_grad():
|
||
output = model.generate(
|
||
input_ids=inputs,
|
||
max_new_tokens=1000,
|
||
temperature=0.7,
|
||
top_p=0.9,
|
||
do_sample=True,
|
||
)
|
||
|
||
print(tokenizer.decode(
|
||
output[0][inputs.shape[1]:],
|
||
skip_special_tokens=True
|
||
))
|
||
|
||
```
|
||
|
||
### Ollama
|
||
|
||
```bash
|
||
# Recommended — best speed/quality balance
|
||
ollama run hf.co/EphAsad/Atem-0.6B:Q4_K_M
|
||
|
||
# Higher quality
|
||
ollama run hf.co/EphAsad/Atem-0.6B:Q5_K_M
|
||
|
||
# Near-lossless
|
||
ollama run hf.co/EphAsad/Atem-0.6B:Q8_0
|
||
|
||
```
|
||
|
||
### llama.cpp
|
||
|
||
```bash
|
||
llama-server -hf EphAsad/Atem-0.6B:Q4_K_M
|
||
|
||
```
|
||
|
||
### System Prompt
|
||
|
||
Atem-0.6B's identity is baked into the chat template and activates automatically when no system message is provided. For manual override:
|
||
|
||
```
|
||
You are Atem, a precise and analytical reasoning assistant. You approach
|
||
every problem methodically — identifying core concepts, reasoning step by
|
||
step, and arriving at well-supported conclusions. You show your thinking
|
||
clearly and are thorough, direct, and intellectually honest.
|
||
|
||
```
|
||
|
||
### Available Files
|
||
|
||
| File | Size | Description |
|
||
| --------------------------- | ---------- | ----------------------------------- |
|
||
| `model.safetensors` | ~1.2 GB | Full bfloat16 merged weights |
|
||
| `Atem-0.6b.Q4_K_M.gguf` | ~397 MB | 4-bit quantised — recommended |
|
||
| `Atem-0.6b.Q5_K_M.gguf` | ~444 MB | 5-bit quantised |
|
||
| `Atem-0.6b.Q8_0.gguf` | ~700 MB | 8-bit quantised — near-lossless |
|
||
|
||
---
|
||
|
||
## Known Limitations
|
||
|
||
**No thinking traces (Stage 1 by design).** Think tags were stripped from all training data for Stage 1, and Qwen3's native exposed reasoning is suppressed. The model does not produce extended `<think>` content. As shown in the qualitative evaluation above, this measurably costs accuracy on multi-step analytical and mathematical problems relative to the base model's default thinking-enabled behavior — Stage 2 (Atem-Savant-0.6B, in progress) exists specifically to recover this.
|
||
|
||
**Smaller capability ceiling than larger Atem models.** At 0.6B parameters, this is the smallest model in the Atem family. Treat it as a fast, low-footprint option rather than a reasoning-depth flagship.
|
||
|
||
**Mathematical precision on complex problems.** On multi-step calculations, the model may make arithmetic or counting errors without a scratchpad to catch them — verified directly in qualitative testing (e.g., miscounting combinatorial outcomes). Answers to high-stakes mathematical problems should be independently verified.
|
||
|
||
**Eval precision asymmetry.** The benchmark comparison above evaluated the base model in 4-bit and Atem-0.6B in bfloat16 — see the Evaluation section for details. A precision-matched re-run would give a cleaner comparison.
|
||
|
||
---
|
||
|
||
## Roadmap
|
||
|
||
Atem-0.6B establishes the Stage 1 foundation for the Qwen3-based branch of the Atem family. Planned next steps:
|
||
|
||
- **Stage 2 (in progress):** Atem-Savant-0.6B — LoRA SFT on curated chain-of-thought data (~90% think-trace records, ~10% no-think) using OpenR1-Math, Kimi-K2.5, DeepSeek-V4-Pro-Reasoning, OpenCodeReasoning, and trace-inversion datasets, to recover multi-step reasoning depth on top of Stage 1's direct-answer foundation
|
||
- **Extended benchmarks:** MMLU, BBH, IFEval post-Stage 2
|
||
- **Precision-matched re-benchmark:** re-run base vs Atem comparison under identical 4-bit (or identical bf16) conditions
|
||
|
||
---
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@misc{atem_06b_2026,
|
||
author = {Asad, Zain},
|
||
title = {Atem-0.6B: A 0.6B Direct-Reasoning Model via
|
||
Stage 1 SFT on Qwen3},
|
||
year = {2026},
|
||
publisher = {HuggingFace},
|
||
howpublished = {\url{https://huggingface.co/EphAsad/Atem-0.6B}},
|
||
}
|
||
|
||
```
|
||
|
||
---
|
||
|
||
## License
|
||
|
||
Released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0), consistent with the base model Qwen/Qwen3-0.6B.
|
||
|
||
---
|
||
|
||
Built independently by Zain Asad - [EphAsad](https://huggingface.co/EphAsad) |