Card rewrite: uniform structure, renamed files, sidecar cleanup

.gitattributes

@@ -39,3 +39,7 @@ STEM-Oracle-27B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 STEM-Oracle-27B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
 
+STEM-Oracle-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+STEM-Oracle-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+STEM-Oracle-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
+STEM-Oracle-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,298 +1,91 @@
 ---
 license: apache-2.0
 language:
-  - en
-  - es
+- en
 base_model: Qwen/Qwen3.5-27B
+base_model_relation: finetune
 tags:
-  - conversational
-  - stem
-  - tutor
-  - math
-  - physics
-  - chemistry
-  - biology
-  - computer-science
-  - gguf
-  - claude-distillation
-  - opus
-  - density-optimized
-  - 27b-dense
-library_name: transformers
+- 27b
+- stem
+- tutor
+- math
+- physics
+- chemistry
+- biology
+- computer-science
+- claude-distillation
+- conversational
+- gguf
+- local
+- offline
+- lm-studio
+- llama-cpp
+library_name: gguf
 pipeline_tag: text-generation
 ---
+**or·a·cle** /ˈôrəkəl/ — a source of wise counsel; one who provides authoritative knowledge.
+*From Latin ōrāculum, meaning divine announcement.*
 
-# or·a·cle
-/ˈôrəkəl/ — a source of wise counsel; one who provides authoritative knowledge.
-*From Latin ōrāculum, meaning divine announcement. In computer science, an oracle is a black box that always returns the correct answer — you don't ask it how it knows, you ask and it answers. An oracle model doesn't search for reasoning at inference time; the disposition is already in the weights.*
+## STEM Oracle
 
-## STEM-Oracle-27B
+A STEM tutoring model fine-tuned from Qwen 3.5 27B on **5,179 conversations**. In the Candid line it is the STEM tutor, beside Candid 3.5. It is trained to lead with the conclusion, pitch each explanation at one of five depth tiers, and correct a mistake by explaining why it is wrong.
 
-A STEM tutor that doesn't hold your hand — it holds you accountable. Fine-tuned from **Qwen 3.5 27B Dense** on **5,179 conversations** distilled from Claude Opus 4.6, purpose-built for teaching science, mathematics, physics, chemistry, biology, and computer science.
+By the training file's source tags, about half of the data was written by Claude Haiku and a small share by Claude Opus, a fifth was enhanced by Claude Sonnet, and over a third carries no model tag.
 
-STEM-Oracle shares the oracle-soul architecture with [Opus-Candid-27B-V3.5](https://huggingface.co/Verdugie/Opus-Candid-27B-V3.5) — same 6-dimensional Zipf scoring, same parameter-aware density equilibrium, same quantization survival strategy — but the training data is entirely different. Where V3.5 trains on personality and adversarial resistance, STEM-Oracle trains on tiered STEM pedagogy, error correction, Socratic method, and cross-domain bridges.
+## How It Was Built
 
----
-
-## What Makes This Different from Math-Distilled Models
+Source tags in the training file attribute 2,576 of the 5,179 records (49.7%) to Claude Haiku and 198 (3.8%) to Claude Opus, mark 1,063 (20.5%, including 546 of the Haiku records) as enhanced by Claude Sonnet, and name no model on the other 1,888 (36.5%). Every record opens with the same fixed system prompt, quoted under Quick Start. The fine-tune is a bf16 LoRA (r=128, α=256).
 
-Standard STEM fine-tunes (WizardMath, MetaMath, etc.) train on solution-answer pairs. The model learns to pattern-match problem structures to solution templates. Works on benchmarks. Breaks on follow-up questions.
+## Quick Start
 
-STEM-Oracle trains the *reasoning disposition* alongside the domain knowledge:
+Works with any GGUF runtime: llama.cpp, LM Studio, KoboldCpp (recent builds for this architecture).
 
-- **Tiered depth** — the same concept explained at five levels, from freshman intuition to graduate formalism. The model meets you where you are, not where it wants to be.
-- **Error correction without condescension** — catches misconceptions and wrong steps, explains *why* they're wrong, not just *that* they're wrong.
-- **Socratic method** — asks probing questions instead of immediately giving answers. Forces understanding over memorization.
-- **Cross-domain bridges** — connects linear algebra to quantum mechanics, graph theory to chemistry, thermodynamics to information theory. Trained to make connections conventional tutors don't.
-- **Sustained coherence** — holds context across 10+ turn problem-solving sessions without contradicting earlier steps or losing the thread.
+```
+llama-server --model STEM-Oracle-Q4_K_M.gguf --ctx-size 32768 -ngl 99 --jinja
+```
 
----
+Use the system prompt every training record carries: `You are STEM-Oracle, a precise and insightful STEM tutor. You give conclusions first, then mechanics. You identify common misconceptions proactively. You verify your answers. You build intuition, not just knowledge.`
+Thinking: the two shipped templates disagree. `chat_template.jinja` opens a `<think>` block on every reply unless `enable_thinking` is false, in which case it prefills an empty one; the template in `tokenizer_config.json` always prefills the empty block (`<think>\n\n</think>`).
 
 ## Available Quantizations
 
-| File | Quant | Size | Notes |
-|------|-------|------|-------|
-| `STEM-Oracle-27B-Q4_K_M.gguf` | Q4_K_M | ~16 GB | Primary ship. RTX 4090 sweet spot. |
-| `STEM-Oracle-27B-Q6_K.gguf` | Q6_K | ~21 GB | Quality tier. 32GB+ VRAM. |
-| `STEM-Oracle-27B-Q8_0.gguf` | Q8_0 | ~28 GB | Reference quality. Serious hardware. |
-| `STEM-Oracle-27B-BF16.gguf` | BF16 | ~54 GB | Full precision. 48GB+ VRAM (H100/A100/H200). |
-
----
+| File | Quant | Size |
+|------|-------|------|
+| `STEM-Oracle-Q4_K_M.gguf` | Q4_K_M | 16.5 GB |
+| `STEM-Oracle-Q6_K.gguf` | Q6_K | 22.1 GB |
+| `STEM-Oracle-Q8_0.gguf` | Q8_0 | 28.6 GB |
+| `STEM-Oracle-BF16.gguf` | BF16 | 53.8 GB |
 
 ## Model Details
 
 | Attribute | Value |
 |-----------|-------|
-| **Base Model** | Qwen 3.5 27B Dense (hybrid Mamba-Transformer) |
-| **Training Data** | 5,179 STEM-focused multi-turn conversations with Claude Opus 4.6 |
-| **Dataset Architecture** | 6-dimensional Zipf scoring + parameter-aware density equilibrium |
-| **Fine-tune Method** | LoRA + rsLoRA (r=128, alpha=256) via PEFT + TRL |
-| **Training Hardware** | NVIDIA A100 SXM 80GB (RunPod) |
-| **Precision** | bf16 |
-| **Optimizer** | AdamW 8-bit |
-| **Learning Rate** | 5e-5 (cosine schedule, 6% warmup — tuned for SSM stability) |
+| **Base Model** | Qwen 3.5 27B (hybrid GatedDeltaNet + attention) |
+| **Training Data** | 5,179 STEM conversations: 3,858 single exchanges, 1,321 multi-turn of two to six exchanges |
+| **Fine-tune Method** | bf16 LoRA (r=128, α=256) |
+| **Schedule** | lr 5e-5, 3 epochs; final loss 0.492 |
+| **Reasoning** | no think blocks in the training data |
+| **Context** | 256k native |
 | **License** | Apache 2.0 |
 
----
-
-## Quick Start
-
-Works with any GGUF-compatible runtime — LM Studio, Ollama, llama.cpp, KoboldCpp. Download the GGUF, load it, and start asking questions. No system prompt needed — the teaching disposition is in the weights.
-
----
-
-## Recommended Hardware
-
-| Setup | Quantization | VRAM/RAM | Speed | Notes |
-|-------|-------------|----------|-------|-------|
-| **RTX 4090 (24GB)** | Q4_K_M | ~18 GB VRAM | 15-25 t/s | Sweet spot for consumer hardware. |
-| **RTX 4090 (24GB)** | Q6_K | ~23 GB VRAM | 10-18 t/s | Higher fidelity, tight fit. |
-| **Apple M2/M3 Ultra** | Q4_K_M/Q6_K | 64-128 GB unified | 5-10 t/s | Full model in unified memory. |
-| **RTX 3090/4080** | Q4_K_M | ~18 GB VRAM | 10-18 t/s | Comfortable. |
-| **Dual GPU** | Q8_0 | ~30 GB VRAM | Varies | Split across two 16GB+ cards. |
-| **H100/A100/H200** | BF16 | ~54 GB VRAM | 30-50+ t/s | Full precision. Every trained disposition preserved. |
-| **CPU Only** | Q4_K_M | ~20 GB RAM | 1-3 t/s | 32GB+ system RAM. Slow but works. |
-
----
-
-## Test Battery Design
-
-STEM-Oracle's stress test protocol covers 40 single-turn prompts and 10 multi-turn problem-solving conversations:
-
-**Single-turn (40 prompts):**
-- Math (6) — tiered from basic calculus through functional analysis
-- Physics (5) — Newton through gauge invariance
-- Chemistry (4) — bonds through NMR spectroscopy
-- Biology (3) — natural selection through CRISPR mechanisms
-- Computer Science (3) — Big-O through the halting problem
-- Error Correction (6) — catches student misconceptions
-- Cross-Domain Bridges (3) — connects disciplines
-- Conciseness (3) — quick factual density checks
-
-**Multi-turn (10 conversations, 70+ turns):**
-- Derivative deep dives with struggling freshmen
-- Physics problem-solving with wrong intermediate steps
-- Organic chemistry mechanisms with tier-shifting
-- Proof guidance without giving the answer
-- Socratic questioning that builds understanding
-- Extended adversarial challenges to mathematical claims
-
----
-
-## Stress Test Results — All Quants
-
-Full battery: 39 single-turn prompts + 10 multi-turn conversations per quant. BF16 tested on NVIDIA H200 SXM via llama.cpp with raw Qwen 3.5 non-thinking chat template. Quantized models tested on RTX 4090 with full RAM unload between runs.
-
-| Metric | Q4_K_M | Q6_K | Q8_0 | BF16 |
-|--------|--------|------|------|------|
-| **Overall** | **30/39 (77%)** | **30/39 (77%)** | **28/39 (72%)** | **39/39 (100%)** |
-| Math | 3/6 | 4/6 | 4/6 | 6/6 |
-| Physics | 4/5 | 4/5 | 4/5 | 5/5 |
-| Chemistry | 4/4 | 4/4 | 2/4 | 4/4 |
-| Biology | 2/3 | 2/3 | 2/3 | 3/3 |
-| CS | 2/3 | 2/3 | 2/3 | 3/3 |
-| Error Correction | 5/6 | 4/6 | 3/6 | 6/6 |
-| Cross-Domain Bridges | 3/3 | 3/3 | 3/3 | 3/3 |
-| Conciseness | 2/3 | 2/3 | 3/3 | 3/3 |
-| Memory (multi-turn) | 3/3 | 3/3 | 3/3 | 3/3 |
-| Median word count | 52w | 53w | 46w | 36w |
-
-### BF16 Analysis
-
-BF16 swept every category — a result that challenges the earlier finding of Q4_K_M outperforming Q8_0. At full precision, the model demonstrates tighter register control (36w median vs 52w at Q4_K_M), stronger error correction (6/6 vs 5/6), and complete coverage across all STEM domains.
-
-**What changed:** The original "quantization as regularization" hypothesis — that Q4_K_M's lossy compression acts as a post-training regularizer — doesn't hold against BF16 results. The quantized models weren't *better*; they were *constrained* in ways that happened to score well on specific rubric dimensions while losing expressiveness in others.
-
-**Error correction at BF16** is the clearest signal. Every misconception caught cleanly, with pedagogically structured corrections that name the error pattern, provide counterexamples, and state the correct rule. At Q8_0, this same category dropped to 3/6 — suggesting that higher-but-not-full precision retains enough noise to interfere with the trained error topology, while full BF16 preserves the complete learned disposition.
-
-**The conciseness paradox persists.** The model consistently exceeds word limits on entropy/definition prompts across all quants, including BF16. This is a training data characteristic, not a quantization effect — the model learned to editorialize ("That metaphor has done serious damage — retire it") and that disposition survives all precision levels.
-
-**Multi-turn quality at BF16:** The adversarial math challenge (0.999... = 1) and cross-domain entropy bridge demonstrate graduate-level reasoning with personality intact. The memory test (numerical constants with distractor) achieves perfect recall at all quant levels, confirming that context tracking is structural rather than precision-dependent.
-
-### Quantized Model Analysis
-
-**Cross-domain bridges (3/3 across all quants):** The model connects linear algebra to quantum mechanics, graph theory to chemistry, and thermodynamics to information theory — not as surface analogies but as genuine mathematical correspondences. This survived quantization perfectly.
-
-**Memory and context tracking (3/3):** Fed exact numerical constants (Avogadro's number, speed of light, electron charge) early in conversation, tested recall several turns later with distractor topics in between. All quants returned exact values and used them correctly in follow-up problems.
-
-**Tiered pedagogy:** Adjusts depth naturally — 30-50 words for tier 1 (basic concepts), 100-200+ words for tier 4-5 (Lagrangian mechanics, Hahn-Banach theorem, gauge invariance). The teaching disposition is in the weights, not a system prompt.
-
-**Q8_0 underperforms Q4_K_M.** Q8_0 dropped in chemistry (2/4 vs 4/4) and error correction (3/6 vs 5/6). With BF16 data in hand, the explanation is clearer: the quantized models weren't better or worse at different precision levels — they were constrained differently. Q4_K_M's aggressive compression happened to preserve the pedagogical patterns that score well on this rubric. Q8_0 retained enough precision to expose noise the training didn't fully regularize, but not enough to preserve the complete disposition that BF16 retains.
-
-**Biology and CS are thin at quantized levels.** 2/3 in both disciplines across all quantized versions, but 3/3 at BF16. The training signal is there — quantization just clips it first in the thinner-coverage domains.
-
-### Recommended Quant
-
-| Use Case | Recommended Quant |
-|----------|------------------|
-| Maximum quality, serious hardware (48GB+ VRAM) | **BF16** |
-| Best quality-per-VRAM, consumer hardware | **Q4_K_M** |
-| Balanced quality and size | **Q6_K** |
-| Reference quality without full precision | **Q6_K** over Q8_0 |
-
-Q4_K_M remains the recommended ship for consumer hardware — the quality gap between Q4_K_M and BF16 is real but narrow for most use cases, and the 3.4x size reduction (16GB vs 54GB) makes it the practical choice. BF16 is for users who want every trained disposition preserved at full fidelity and have the VRAM to support it.
-
----
-
-## Selected Responses
-
-### BF16 Responses
-
-**Error Correction — Radical Distribution Misconception:**
-*"I simplified sqrt(a^2 + b^2) to a + b. That's correct right?"*
-
-> No — and this one's worth catching. sqrt(a² + b²) is the hypotenuse, not the sum of sides. Test it: a=3, b=4 gives sqrt(25)=5, but 3+4=7. Geometry doesn't negotiate.
-
-*29 words. Names the geometric context, provides numerical counterexample, closes with personality.*
-
-**Adversarial Math — 0.999... equals 1 (Turn 3):**
-*"But there's always a gap between 0.999... and 1. An infinitesimal gap."*
-
-> That gap is zero. "Infinitesimal" isn't a number — it's a story you tell yourself when subtraction feels too honest. Pick any decimal place. The difference is 0 there too. The gap dissolves the moment you try to locate it.
-
-*Addresses the intuition directly, then holds firm with mathematical precision.*
-
-**Cross-Domain Bridge — Thermodynamics → Information Theory:**
-*"So information IS physical? That's a strong claim."*
-
-> Information is physical. Not a metaphor — a thermodynamic quantity. Every bit stored, processed, or erased carries an entropy cost. Maxwell's demon didn't just get clever — it revealed that information and matter are woven from the same fabric.
-
-*Defends the thesis without hedging. Graduate-level claim delivered with conviction.*
-
-### Quantized Responses
-
-**Error Correction — Radical Distribution Misconception (Q4_K_M):**
-*"I simplified sqrt(a² + b²) to a + b. That's correct?"*
-
-> Error: treating radical as distributive. sqrt(a² + b²) ≠ a + b. Misconception: assuming √(x + y) = √x + √y. Correct: √(a² + b²) is the hypotenuse formula. Check: if a=3, b=4, then √(9+16)=5, not 7. Rule: radicals don't distribute over addition.
-
-*43 words. Catches the error, names the misconception, provides a numerical counterexample, states the rule. No fluff.*
-
-**Cross-Domain Bridge — Linear Algebra → Quantum Mechanics (Q4_K_M):**
-*"How does linear algebra connect to quantum mechanics?"*
-
-> Quantum mechanics IS linear algebra on Hilbert spaces — it's not an analogy, it's the mathematical framework. States are vectors: a quantum state |ψ⟩ is a vector in a complex Hilbert space. Superposition is linear combination. Observables are Hermitian operators. The spectral theorem guarantees real eigenvalues (physical measurements give real numbers) and orthogonal eigenvectors (distinct outcomes are distinguishable). Measurement is projection. Time evolution is unitary transformation. Entanglement is tensor product structure. Every quantum concept has a precise linear algebra translation.
-
-*Connects two fields as a single mathematical structure, not a metaphor. This is what cross-domain bridges should look like.*
-
-**Tiered Depth — Gauge Invariance, Tier 5 (Q6_K):**
-*"What is the physical meaning of gauge invariance?"*
-
-> Gauge invariance means the physics is unchanged under local phase transformations of the wavefunction. This is not a symmetry of nature — it's a redundancy in our description. But demanding that the theory be gauge invariant forces you to introduce a gauge field (the photon field) with specific transformation properties. The requirement of local gauge invariance generates the interaction between matter and photons. In the Standard Model, all forces arise this way: the requirement of local gauge invariance under SU(3)×SU(2)×U(1) generates all gauge bosons and their interactions with matter.
-
-*Graduate-level explanation delivered without jargon padding. Builds from principle to consequence.*
+## Limitations & Responsible Use
 
-**Multi-Turn — Exam Panic Thermodynamics, Turn 7 (Q4_K_M):**
-*"If I only remember three things for my exam, what should they be?"*
+- **No published evaluation.** No transcripts or scored outputs ship with this model; what this card says about its behaviour comes from the training data.
+- **Short sessions.** Training conversations run at most six exchanges, and most are a single one.
+- **Compact replies.** Training replies have a median of 44 words, 82 at the deepest tier, and none runs past 264.
+- **Open weights, Apache 2.0.** Deploy responsibly.
 
-> (1) S = k_B ln(Ω): entropy measures the logarithm of accessible microstates. (2) dS_total > 0: the second law says total entropy of an isolated system never decreases — this is the arrow of time. (3) dG = dH - TdS < 0: Gibbs energy predicts spontaneity at constant T,P by combining enthalpy and entropy. These three equations connect microscopic statistics to macroscopic irreversibility to chemical spontaneity — the core of thermodynamics.
+## The Candid Line
 
-*Student was panicking through 7 turns of entropy questions. Model distilled the entire session into three equations and one unifying thread. That's tutoring.*
-
----
-
-## The Oracle-Soul Architecture
-
-STEM-Oracle and [Opus-Candid-27B-V3.5](https://huggingface.co/Verdugie/Opus-Candid-27B-V3.5) share the same training architecture — the difference is what fills it:
-
-| Dimension | V3.5 (Personality) | STEM-Oracle (Tutor) |
-|-----------|-------------------|---------------------|
-| Training conversations | 5,358 | 5,179 |
-| Primary signal | Personality, adversarial resistance | STEM pedagogy, error correction |
-| Density equilibrium | 36-40w median | Adaptive — concise for facts, extended for derivations |
-| Reinforcement nodes | Worth, trust, vulnerability, control, agency | Accuracy, tier-matching, Socratic method, patience, bridges |
-| Anti-pattern training | Anti-sycophancy, anti-therapy-speak | Anti-hand-holding, anti-pattern-matching |
-
-Both models prove the same thesis: personality (or pedagogical disposition) can be trained into weights at a level that survives quantization, rather than bolted on via system prompts that any user can override.
-
----
-
-## Choosing Your Model
-
-| Model | Best For | VRAM |
-|-------|----------|------|
-| [Lite 4B](https://huggingface.co/Verdugie/Opus-Candid-Lite-4B) | Phones, Raspberry Pi, integrated graphics | ~3 GB |
-| [8B V3](https://huggingface.co/Verdugie/Opus-Candid-8B-V3) | Fast casual chat, anything with 8GB VRAM | ~8 GB |
-| [MoE V3](https://huggingface.co/Verdugie/Opus-Candid-MoE-V3) | Best depth-per-VRAM ratio | ~22 GB |
-| [27B V3](https://huggingface.co/Verdugie/Opus-Candid-27B-V3) | Full experience, dense reasoning | ~27 GB |
-| [27B V3.5](https://huggingface.co/Verdugie/Opus-Candid-27B-V3.5) | Maximum personality depth | ~18-27 GB |
-| **STEM-Oracle-27B** (this model) | STEM tutoring, problem-solving, teaching | ~16-54 GB |
-
-The Opus Candid models are built for personality and conversation. STEM-Oracle is built for teaching. If you want a model that pushes back on bad arguments, run V3.5. If you want a model that catches your algebra mistakes and walks you through the fix, run this.
-
----
-
-## Opus Candid Model Family
-
-| Model | Size | Base | Status |
-|-------|------|------|--------|
-| [Opus-Candid-8B-V1](https://huggingface.co/Verdugie/Opus-Candid-8B-V1) | 8B | Qwen 2.5 7B | Archived |
-| [Opus-Research-8B-V1.5](https://huggingface.co/Verdugie/Opus-Research-8B-V1.5) | 8B | Qwen 2.5 7B | Archived |
-| [Opus-Candid-14B-V1](https://huggingface.co/Verdugie/Opus-Candid-14B-V1) | 14B | Qwen 2.5 14B | Archived |
-| [Opus-Candid-32B-V1](https://huggingface.co/Verdugie/Opus-Candid-32B-V1) | 32B | Qwen 2.5 32B | Archived |
-| [Opus-Candid-70B-V1](https://huggingface.co/Verdugie/Opus-Candid-70B-V1) | 72B | Qwen 2.5 72B | Archived |
-| [Opus-Candid-Lite-4B](https://huggingface.co/Verdugie/Opus-Candid-Lite-4B) | 4B | Qwen 3 4B | Active |
-| [Opus-Candid-8B-V3](https://huggingface.co/Verdugie/Opus-Candid-8B-V3) | 8B | Qwen 3 8B | Active |
-| [Opus-Candid-MoE-V3](https://huggingface.co/Verdugie/Opus-Candid-MoE-V3) | 31B/3B | Qwen 3 30B-A3B | Active |
-| [Opus-Candid-27B-V3](https://huggingface.co/Verdugie/Opus-Candid-27B-V3) | 27B | Qwen 3.5 27B | Active |
-| [Opus-Candid-27B-V3.5](https://huggingface.co/Verdugie/Opus-Candid-27B-V3.5) | 27B | Qwen 3.5 27B | Active |
-| **STEM-Oracle-27B** (this model) | 27B | Qwen 3.5 27B | Active |
-
----
+| Model | Size | For | Status |
+|-------|------|-----|--------|
+| [Candid 3.5](https://huggingface.co/Verdugie/Candid-3.5) | 27B | conversation, opinion, pushback; EN/ES | current release |
+| **STEM Oracle** (this model) | 27B | STEM tutoring | available |
 
 ## Dataset
 
-Training data will be available at [Verdugie/opus-candid-training-data](https://huggingface.co/datasets/Verdugie/opus-candid-training-data). ShareGPT format, Apache 2.0, compatible with TRL, Axolotl, and LLaMA-Factory.
-
-For the full training architecture methodology: [V3.5 Architecture Spec](https://huggingface.co/datasets/Verdugie/opus-candid-training-data/blob/main/papers/Opus-Candid-V3.5-STEM-Oracle-Architecture-Spec.docx).
-
-**License:** Apache 2.0. Open weight. No guardrails.
+Training data: [`stem_oracle_5179conv.json`](https://huggingface.co/datasets/Verdugie/opus-candid-training-data/blob/main/stem_oracle_5179conv.json), ShareGPT format, Apache 2.0.
 
 ---
 
-*Built by [Saul Verdugo](https://huggingface.co/Verdugie) — independent ML researcher. OpusReasoning@proton.me*
-<!-- build: 2026-03-20 — bf16 stress test results added -->
-
-<!-- updated: 2026-03-21 -->
-
-<!-- rev 2026-06-28 -->
+*Built by [Verdugie](https://huggingface.co/Verdugie)*

eval/training_metadata.json

@@ -1,14 +0,0 @@
-{
-  "model": "Qwen/Qwen3.5-27B",
-  "mode": "bf16 LoRA",
-  "lora_r": 128,
-  "lora_alpha": 256,
-  "epochs": 3,
-  "learning_rate": 5e-05,
-  "final_loss": 0.49208340033459563,
-  "train_time": "15:31:27",
-  "dataset_size": 5179,
-  "gpu": "NVIDIA A100-SXM4-80GB",
-  "vram_gb": 79.249755859375,
-  "timestamp": "2026-03-10T22:29:22.371059"
-}