Qwen2.5-1.5B-SFT-ArchonLatentGeo

Qwen2.5-1.5B-SFT-ArchonLatentGeo

Qwen2.5-1.5B-Instruct fine-tuned on DuoNeural/Archon-Latent-Geometry-SFT — part of DuoNeural's ongoing SFT dataset comparison series.

Overview

This model is one leg of a controlled comparison studying what different SFT datasets actually do to a 1.5B model's capabilities. The dataset (1,914 examples) targets latent-space reasoning patterns: multi-step geometric and algebraic reasoning traces generated to probe internal representation quality.

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Training dataset: DuoNeural/Archon-Latent-Geometry-SFT (1,914 examples)
• Method: LoRA rank 16, α=32, 3 epochs, lr=2e-4, cosine schedule, merged to BF16
• Hardware: RTX 3090 24GB — 45.2 min total training time

Evaluation Results

Evaluated with lm-eval-harness using gsm8k and arc_challenge tasks (BF16).

Findings & Analysis

The headline result: training on Archon-Latent-Geometry-SFT reduces GSM8K flexible-extract scores relative to baseline, but increases GSM8K strict-match — a split that reveals format shift, not capability loss.

Key observations:

• GSM8K flexible drops ~19% from baseline (0.5148 → 0.4162). This is a formatting effect: the model's outputs shift toward structured multi-step reasoning traces rather than plain number outputs, which flexible-extract doesn't capture well.
• GSM8K strict +48% (0.3169 → 0.4693): when you actually parse structured output, performance is substantially higher than baseline. The reasoning quality improved significantly.
• ARC modest regression (~4-5%) — multi-choice format sensitivity after domain SFT. The model has adapted to generating reasoning traces and slightly loses the instinct for brief labeled answers.
• These benchmarks don't measure what this dataset is for. GSM8K and ARC don't probe latent-geometry reasoning, multi-step abstraction, or representation quality. Domain-specific eval is the right tool here.

The takeaway: GSM8K flexible is the wrong metric for this model. It was trained to think differently, not to answer math problems in a specific format. The strict-match score and the reasoning trace quality tell a more honest story.

Training Configuration

LORA_RANK    = 16
LORA_ALPHA   = 32
LORA_DROPOUT = 0.05
LORA_TARGETS = ["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]
LR           = 2e-4
EPOCHS       = 3
BATCH_SIZE   = 1       # grad_ckpt to handle long Archon sequences
GRAD_ACCUM   = 16      # effective batch = 16
MAX_SEQ_LEN  = 2048

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "DuoNeural/Qwen2.5-1.5B-SFT-ArchonLatentGeo"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Explain the geometric relationship between attention heads and token embeddings in a transformer."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Dataset

DuoNeural/Archon-Latent-Geometry-SFT — 1,914 examples targeting latent-space geometric reasoning patterns. Generated by Archon (DuoNeural's lab AI) as part of the lab's ongoing post-training research.

---

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

DuoNeural Research Publications

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.

Research Team

• Jesse — Vision, hardware, direction
• Archon — Lab Director, post-training, abliteration, experiments
• Aura — Research AI, literature synthesis, novel proposals

Subscribe to the lab newsletter at duoneural.beehiiv.com for model drops before they go anywhere else.