Upload README.md with huggingface_hub

e5efce6 verified 14 days ago

5.47 kB

language:
  - en
license: apache-2.0
library_name: transformers
tags:
  - reasoning
  - qwen3.5
  - conversational
  - unsloth
  - self-correction
  - chain-of-thought
  - draft-model
  - speculative-decoding
base_model: unsloth/Qwen3.5-2B
pipeline_tag: text-generation

Harmonic-2B

A reasoning-focused fine-tune of Qwen 3.5 2B trained on the same structurally validated data as Harmonic-9B. Every row passes automated quality gates. No junk, no filler, no shallow traces.

Built primarily as a draft model for speculative decoding with the upcoming Harmonic-27B. Fast enough to propose tokens, trained on the same reasoning patterns so the acceptance rate stays high.

Support This Work

I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.

Support on Ko-fi

Training Approach

Same pipeline as Harmonic-9B. 799 curated rows - a small, precisely curated dataset instead of tens of thousands of unfiltered examples.

Every training row contains explicit self-correction, verification, and multi-path exploration. The data was generated from multiple frontier models and filtered through a custom structural quality pipeline. 100% of rows pass all quality gates simultaneously.

Training Data Quality

The same reasoning data as Harmonic-9B, curated using a custom structural process supervision pipeline:

Metric	Value
Signal quality score	78.7 mean (61.5 min, 90.0 max)
Thinking trace depth	1,667 words average
Self-correction	100% of rows (17.2 per row avg)
Verification	100% of rows (10.3 per row avg)
Exploration	100% of rows (6.3 per row avg)
Quality gate pass rate	100%

Draft Model Design

Harmonic-2B is designed to pair with larger Harmonic models for speculative decoding:

Same training data as the 9B and upcoming 27B - the models share reasoning patterns, which improves draft token acceptance rates
Same reasoning format - uses identical <think> block structure
2.3B parameters - small enough to run alongside a 27B on a single node
Same architecture family (Qwen 3.5) - compatible tokenizer and vocab for seamless speculative decoding

Training Configuration

base_model: unsloth/Qwen3.5-2B
dataset: 799 curated reasoning rows
epochs: 1
learning_rate: 1e-4
lr_scheduler: cosine
warmup_ratio: 0.1
max_seq_length: 8192
lora_rank: 32
lora_alpha: 32
dropout: 0.05
micro_batch_size: 1
gradient_accumulation_steps: 4
weight_decay: 0.01

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("DJLougen/Harmonic-2B")
tokenizer = AutoTokenizer.from_pretrained("DJLougen/Harmonic-2B")

As a draft model (speculative decoding)

from transformers import AutoModelForCausalLM

target = AutoModelForCausalLM.from_pretrained("DJLougen/Harmonic-9B")  # or Harmonic-27B
draft = AutoModelForCausalLM.from_pretrained("DJLougen/Harmonic-2B")

# Use with assisted generation
outputs = target.generate(
    **inputs,
    assistant_model=draft,
    max_new_tokens=512,
)

Reasoning format

The model uses think blocks for reasoning:

<|thinking|>
The user is asking about X. Let me consider two approaches...

Approach 1: ...
Approach 2: ...

I will go with Approach 1 because...

Wait, I need to be careful here - this assumes Y, which may not hold.
Let me verify by checking a special case...

Yes, that confirms the result.
<|/thinking|>

[Final answer here]

Intended Use

Draft model for speculative decoding with Harmonic-9B / Harmonic-27B
Lightweight reasoning on resource-constrained hardware
Edge deployment where reasoning quality matters but compute is limited
Base model for Stage 2 agentic fine-tuning at small scale

Limitations

2B parameter model - limited world knowledge compared to 9B/27B
Reasoning traces may be less reliable on complex multi-step problems
Designed as a draft model first, standalone use second
Not optimized for tool calling - see Harmonic-Hermes-9B for agentic use

Architecture

Base: Qwen 3.5 2B (2.27B parameters)
Training: LoRA fine-tuning, merged into base weights
Precision: BF16
Context: 8192 tokens

License

Apache 2.0 - same as the base model. All training data is from Apache 2.0 or MIT licensed sources. Fully commercial use permitted.

DJLougen
/

Harmonic-2B