E4a Solver LoRA β€” Hint-Conditioned Multi-Turn Physics Solver (Final v3)

Part of Adversarial Physics Solver–Verifier (Track 1: Data-Scarce University Physics). Multi-turn hint-conditioned Solver (CohereLabs/tiny-aya-global, Model A) designed to solve Turn-0 physics problems and revise derivations when given self-contained [[CHALLENGE]] directives from the E4b Verifier without repeating mistakes. Extends E3 Turn-0 derivations with harvested valid tiny-aya-global attempts, aya-expanse-8b Alternative Solution 2s, and Turn-1/Turn-2 corrective hint trajectories.

Training Data & Provenance

  • Dataset Repo: vital-ar/physics-solver-e4-adversarial-v3 (758 training examples; strict < 0.80 MiniLM separation vs. Clean held-out splits)
  • Turn Mix & Sources: Turn 0: 623, Turn 1 (1 hint): 100, Turn 2 (2 hints): 35 | Sources: literature_curated: 720, physgym_seed: 7, synthetic_seed_subtopic: 31
  • Domains & Role Split: ELECTROMAGNETISM 138, MECHANICS 421, MODERN_PHYSICS 69, OPTICS 37, RELATIVITY 59, THERMODYNAMICS 34. Data filtering/generation used CohereForAI/aya-expanse-8b exclusively; training & failure harvesting used CohereLabs/tiny-aya-global exclusively.
  • v3 Data Curation: Downgraded fallback 2nd hints to Turn-1, capped completion multiplicity <= 2, >= 400-char clean steps.

Training Configuration

  • Version Strategy (v3): Final regularized v3 run: r=16/alpha=24 Solvers (dropout=0.08), r=8/alpha=16 Verifier (lr=1.0e-4, wd=0.03, 10% val split), short-key/formula-leak pruning, **Derivation:**\n prefix parity, and < 0.80 clean audit.
  • Hyperparameters (AlignTune aligntune==0.1.32): Base CohereLabs/tiny-aya-global (3.35B FP16), LoRA r=16 / alpha=24, LR 0.0001, epochs 5.0/7, steps 240, final train loss 0.788, selected checkpoint checkpoint-96, completion_only_loss=true, seed=42.

Evaluation (Clean cosine < 0.80 Splits)

Clean Split Overall Acc Textbook Acc Sentinel % Degenerate % AuditKIT SHA-256
validation (n=48) 2.08% (1/48) 10.00% (1/10) 89.58% 2.08% 54cbe44db34f8fde
test (n=55) 3.64% (2/55) 11.76% (2/17) 94.55% 1.82% 29b4374863bc2dfb
  • Hard held-out YimengChen/PhysGym (38 problems) stays 0/38 across all runs (see main repo for root-cause analysis and Table 1 pre-filter comparison).

Usage (AlignTune Raw Prompt Format)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "CohereLabs/tiny-aya-global"
adapter_id = "vital-ar/adversarial-physics-solver-e4-adversarial-v3-lora"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id).eval()
prompt = (
    "Solve the following physics problem step by step (in at most 10 logical steps). "
    "State the governing physical equations, perform algebraic substitutions clearly, include "
    "SI units, and end with the final answer on a new line prefixed with ANSWER:.\n\n"
    "PROBLEM:\nA 2 kg block starts from rest on a 5 m frictionless ramp. Find its speed "
    "at the bottom.\n\n"
)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Links & Attribution

Downloads last month
44
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for vital-ar/adversarial-physics-solver-e4-adversarial-v3-lora

Adapter
(393)
this model

Datasets used to train vital-ar/adversarial-physics-solver-e4-adversarial-v3-lora

Collection including vital-ar/adversarial-physics-solver-e4-adversarial-v3-lora